docs: link workspace integration + gitea docs

- api_confirm and api_reject were defined twice on same paths
- FastAPI only registers first definition, causing silent 404s
- Kept api_confirm_engram and api_reject_engram (use _update_correctness)
- Removed duplicate direct DB implementations
- Fixes dashboard confirm/reject buttons not working

RUNBOOK.md

@@ -2,19 +2,32 @@
 
 This is the operational quick-reference for the shipped systemd timers/services and the optional FastAPI dashboard backend.
 
-Repository root (on host): `/root/.openclaw/workspace`
+Repository root (on host): `/root/.openclaw/workspace/second-brain`
+
+Integration + orchestration docs live in the workspace repo:
+
+- `/root/.openclaw/workspace/docs/SECOND_BRAIN.md`
+- `/root/.openclaw/workspace/docs/GITEA.md`
 
 ## Systemd units (cron jobs)
 
-Unit files are shipped in `systemd/` (repo root). Install them into `/etc/systemd/system/` (symlink or copy), then reload:
+Unit files are shipped in `systemd/` (this repo). Install them into `/etc/systemd/system/` (symlink or copy), then reload:
 
 ```bash
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-secondbrain-*.service /etc/systemd/system/
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-secondbrain-*.timer /etc/systemd/system/
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-memory-archive.* /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-*.service /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-*.timer /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-memory-archive.* /etc/systemd/system/
 sudo systemctl daemon-reload
 ```
 
+Optional (verification hardening):
+
+```bash
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-verify-pending.* /etc/systemd/system/
+sudo systemctl daemon-reload
+sudo systemctl enable --now openclaw-secondbrain-verify-pending.timer
+```
+
 Enable timers:
 
 ```bash
@@ -61,7 +74,7 @@ python3 -m pip install -r second-brain/requirements-dashboard.txt
 SECOND_BRAIN_WORKSPACE="/root/.openclaw/workspace/second-brain" python3 second-brain/fastapi_app.py
 ```
 
-Default port is `8501` (same as Streamlit default). Do not run both on the same port.
+Default port is `8501` (same as Streamlit default). You can override via `SECOND_BRAIN_PORT` (or `PORT`) when starting manually.
 
 Endpoint smoke tests:
 

cron_tasks/ingest_transcript_to_db.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+"""
+Ingest OpenClaw session transcript JSONL directly into the Second-Brain DB.
+
+State is tracked with byte offsets per transcript file.
+Sources are configured via workspace/memory/session_sources.json.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, Iterable, List
+from uuid import NAMESPACE_URL, uuid5
+
+WORKSPACE = Path("/root/.openclaw/workspace")
+MEMORY_DIR = WORKSPACE / "memory"
+SOURCES_PATH = MEMORY_DIR / "session_sources.json"
+STATE_PATH = MEMORY_DIR / "session_db_ingest_state.json"
+
+BRAIN_DIR = WORKSPACE / "second-brain"
+DB_PATH = Path(os.environ.get("BRAIN_DB", str(BRAIN_DIR / "data" / "brain.sqlite")))
+
+import sys
+
+sys.path.insert(0, str(BRAIN_DIR))
+from src.engram import Engram, Grounding  # type: ignore
+from src.store import EngramStore  # type: ignore
+
+
+def _now() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _load_json(path: Path, default: Any) -> Any:
+    try:
+        if not path.exists():
+            return default
+        return json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return default
+
+
+def _save_json(path: Path, payload: Any) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
+
+
+def _extract_text(content: Any) -> str:
+    if isinstance(content, str):
+        return content.strip()
+    if isinstance(content, list):
+        parts: List[str] = []
+        for c in content:
+            if isinstance(c, dict) and c.get("type") == "text" and isinstance(c.get("text"), str):
+                parts.append(c["text"])
+            elif isinstance(c, str):
+                parts.append(c)
+        return "\n".join(p.strip() for p in parts if p and p.strip()).strip()
+    if isinstance(content, dict) and isinstance(content.get("text"), str):
+        return content["text"].strip()
+    return str(content).strip()
+
+
+@dataclass
+class Source:
+    label: str
+    transcript_path: Path
+
+
+def _load_sources() -> List[Source]:
+    payload = _load_json(SOURCES_PATH, {"sources": []})
+    sources: List[Source] = []
+    for item in payload.get("sources", []) if isinstance(payload, dict) else []:
+        if not isinstance(item, dict):
+            continue
+        label = str(item.get("label") or "session")
+        p = item.get("path")
+        if not isinstance(p, str) or not p:
+            continue
+        sources.append(Source(label=label, transcript_path=Path(p)))
+    return sources
+
+
+def _iter_new_lines(path: Path, *, start_offset: int) -> Iterable[tuple[int, str]]:
+    with open(path, "rb") as f:
+        f.seek(max(0, int(start_offset)))
+        while True:
+            raw = f.readline()
+            if not raw:
+                break
+            line = raw.decode("utf-8", errors="ignore").strip()
+            if not line:
+                continue
+            yield (f.tell(), line)
+
+
+def run() -> Dict[str, Any]:
+    if not DB_PATH.exists():
+        return {"success": False, "time": _now(), "error": f"db missing: {DB_PATH}"}
+
+    sources = _load_sources()
+    state = _load_json(STATE_PATH, {"offsets": {}})
+    offsets: Dict[str, int] = state.get("offsets", {}) if isinstance(state, dict) else {}
+
+    store = EngramStore(str(DB_PATH))
+    out = {"success": True, "time": _now(), "sources": len(sources), "messages_saved": 0, "messages_skipped": 0, "errors": []}
+
+    for src in sources:
+        try:
+            if not src.transcript_path.exists():
+                continue
+            key = str(src.transcript_path)
+            start = int(offsets.get(key, 0))
+            for new_off, line in _iter_new_lines(src.transcript_path, start_offset=start):
+                offsets[key] = new_off
+                obj = json.loads(line)
+                if not isinstance(obj, dict) or obj.get("type") != "message":
+                    continue
+                msg = obj.get("message") if isinstance(obj.get("message"), dict) else {}
+                role = str(msg.get("role") or "unknown")
+                content = _extract_text(msg.get("content"))
+                if len(content.strip()) < 5:
+                    continue
+                mid = str(obj.get("id") or msg.get("id") or msg.get("messageId") or msg.get("message_id") or "")
+                if not mid:
+                    mid = str(uuid5(NAMESPACE_URL, f"openclaw-transcript:{src.label}:{role}:{content[:200]}"))
+                eid = str(uuid5(NAMESPACE_URL, f"openclaw-transcript:{src.label}:{mid}"))
+                if store.get(eid):
+                    out["messages_skipped"] += 1
+                    continue
+                eg = Engram.create(
+                    content=f"[transcript:{src.label}] [{role}] [{mid}]\n\n{content}"[:4000],
+                    source="session",
+                    tags=["session", "openclaw", "transcript", f"role:{role}"],
+                    session_id=src.label,
+                    confidence=0.55,
+                    grounding=Grounding.ASSUMPTION,
+                )
+                eg.id = uuid5(NAMESPACE_URL, eid)
+                eg.metadata["source"] = "session"
+                eg.metadata["session_id"] = src.label
+                eg.metadata["role"] = role
+                eg.metadata["message_id"] = mid
+                store.save(eg)
+                out["messages_saved"] += 1
+        except Exception as e:
+            out["success"] = False
+            out["errors"].append(f"{src.transcript_path}: {e}")
+
+    _save_json(STATE_PATH, {"offsets": offsets, "updated_at": out["time"]})
+    return out
+
+
+if __name__ == "__main__":
+    print(json.dumps(run(), ensure_ascii=False, indent=2))
+

cron_tasks/ingest_transcript_to_memory.py

@@ -0,0 +1,213 @@
+#!/usr/bin/env python3
+"""
+Tail OpenClaw session transcript JSONL files and append new messages into
+workspace/memory/YYYY-MM-DD.md so the existing ingest_memory pipeline can pick
+them up.
+
+Why: when chat_autosave hooks are missed/aborted, the "memory/*.md -> DB" ingest
+doesn't see the latest conversation. This bridges transcript -> memory.
+
+Safety: read-only access to transcript files; state stored in workspace/memory/.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional
+
+try:
+    from zoneinfo import ZoneInfo  # py3.9+
+except Exception:  # pragma: no cover
+    ZoneInfo = None  # type: ignore
+
+
+WORKSPACE = Path("/root/.openclaw/workspace")
+MEMORY_DIR = WORKSPACE / "memory"
+SOURCES_PATH = MEMORY_DIR / "session_sources.json"
+STATE_PATH = MEMORY_DIR / "session_ingest_state.json"
+
+
+def _local_tz():
+    tz = os.environ.get("TZ") or "Europe/Berlin"
+    if ZoneInfo:
+        try:
+            return ZoneInfo(tz)
+        except Exception:
+            return timezone.utc
+    return timezone.utc
+
+
+def _load_json(path: Path, default: Any) -> Any:
+    try:
+        if not path.exists():
+            return default
+        return json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return default
+
+
+def _save_json(path: Path, payload: Any) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
+
+
+def _iso_to_dt(ts: str) -> Optional[datetime]:
+    try:
+        if ts.endswith("Z"):
+            ts = ts[:-1] + "+00:00"
+        return datetime.fromisoformat(ts)
+    except Exception:
+        return None
+
+
+def _ms_to_dt(ms: int) -> datetime:
+    return datetime.fromtimestamp(ms / 1000, tz=timezone.utc)
+
+
+def _extract_text(content: Any) -> str:
+    if isinstance(content, str):
+        return content.strip()
+    if isinstance(content, list):
+        parts: List[str] = []
+        for c in content:
+            if isinstance(c, dict) and c.get("type") == "text" and isinstance(c.get("text"), str):
+                parts.append(c["text"])
+            elif isinstance(c, str):
+                parts.append(c)
+        return "\n".join(p.strip() for p in parts if p and p.strip()).strip()
+    if isinstance(content, dict):
+        if isinstance(content.get("text"), str):
+            return content["text"].strip()
+    return str(content).strip()
+
+
+@dataclass
+class Source:
+    label: str
+    transcript_path: Path
+
+
+def _load_sources() -> List[Source]:
+    """
+    Sources file format:
+    {
+      "sources": [
+        { "label": "telegram:263887248", "path": "/root/.openclaw/agents/main/sessions/<id>.jsonl" }
+      ]
+    }
+    """
+    payload = _load_json(SOURCES_PATH, {"sources": []})
+    sources: List[Source] = []
+    for item in payload.get("sources", []) if isinstance(payload, dict) else []:
+        if not isinstance(item, dict):
+            continue
+        label = str(item.get("label") or "session")
+        p = item.get("path")
+        if not isinstance(p, str) or not p:
+            continue
+        sources.append(Source(label=label, transcript_path=Path(p)))
+    return sources
+
+
+def _memory_path_for(dt: datetime) -> Path:
+    tz = _local_tz()
+    local = dt.astimezone(tz)
+    MEMORY_DIR.mkdir(parents=True, exist_ok=True)
+    return MEMORY_DIR / f"{local.date().isoformat()}.md"
+
+
+def _append_memory(dt: datetime, label: str, role: str, text: str) -> None:
+    if not text.strip():
+        return
+    tz = _local_tz()
+    local = dt.astimezone(tz)
+    mem = _memory_path_for(dt)
+    if not mem.exists():
+        mem.write_text(f"# {local.date().isoformat()}\n\n", encoding="utf-8")
+
+    header = f"## {local.strftime('%H:%M:%S')} - {label} ({role})"
+    body = text.strip()
+    mem.write_text(mem.read_text(encoding="utf-8") + f"{header}\n\n{body}\n\n", encoding="utf-8")
+
+
+def _iter_new_lines(path: Path, *, start_offset: int) -> Iterable[tuple[int, str]]:
+    with open(path, "rb") as f:
+        f.seek(max(0, int(start_offset)))
+        while True:
+            raw = f.readline()
+            if not raw:
+                break
+            try:
+                line = raw.decode("utf-8", errors="ignore").strip()
+            except Exception:
+                line = ""
+            if not line:
+                continue
+            yield (f.tell(), line)
+
+
+def run() -> Dict[str, Any]:
+    sources = _load_sources()
+    state = _load_json(STATE_PATH, {"offsets": {}})
+    offsets: Dict[str, int] = state.get("offsets", {}) if isinstance(state, dict) else {}
+
+    out = {
+        "success": True,
+        "time": datetime.now(timezone.utc).isoformat(),
+        "sources": len(sources),
+        "messages_appended": 0,
+        "errors": [],
+    }
+
+    for src in sources:
+        try:
+            if not src.transcript_path.exists():
+                continue
+            key = str(src.transcript_path)
+            start = int(offsets.get(key, 0))
+            for new_off, line in _iter_new_lines(src.transcript_path, start_offset=start):
+                try:
+                    obj = json.loads(line)
+                except Exception:
+                    offsets[key] = new_off
+                    continue
+
+                if not isinstance(obj, dict) or obj.get("type") != "message":
+                    offsets[key] = new_off
+                    continue
+
+                msg = obj.get("message") if isinstance(obj.get("message"), dict) else {}
+                role = str(msg.get("role") or "unknown")
+                content = msg.get("content")
+                text = _extract_text(content)
+
+                dt = None
+                if isinstance(msg.get("timestamp"), (int, float)):
+                    dt = _ms_to_dt(int(msg["timestamp"]))
+                elif isinstance(obj.get("timestamp"), str):
+                    dt = _iso_to_dt(obj["timestamp"])
+                if dt is None:
+                    dt = datetime.now(timezone.utc)
+
+                if len(text.strip()) < 5:
+                    offsets[key] = new_off
+                    continue
+
+                _append_memory(dt, src.label, role, text)
+                out["messages_appended"] += 1
+                offsets[key] = new_off
+        except Exception as e:
+            out["success"] = False
+            out["errors"].append(f"{src.transcript_path}: {e}")
+
+    _save_json(STATE_PATH, {"offsets": offsets, "updated_at": out["time"]})
+    return out
+
+
+if __name__ == "__main__":
+    res = run()
+    print(json.dumps(res, ensure_ascii=False, indent=2))

cron_tasks/verify_pending_external.py

@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+"""
+Verify pending (unconfirmed) engrams using lightweight external checks.
+
+Policy (conservative):
+- `openclaw-memory` is treated as internal ground-truth and is auto-confirmed
+  by the review job (see `cron_tasks/review_brain.py` in the workspace runtime).
+- For `source=web`, confirm if the grounded URL responds with HTTP 2xx, reject on
+  4xx/5xx, and keep pending on timeouts/unknown.
+- Reject obvious low-signal placeholders (e.g. session summary stubs).
+"""
+
+import json
+import os
+import sys
+from pathlib import Path
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+
+WORKSPACE = Path(os.environ.get("SECOND_BRAIN_WORKSPACE", "/root/.openclaw/workspace/second-brain"))
+DB_PATH = Path(os.environ.get("BRAIN_DB", str(WORKSPACE / "data" / "brain.sqlite"))).resolve()
+
+sys.path.insert(0, str(WORKSPACE))
+from src.store import EngramStore
+from src.engram import ReviewEntry
+
+OUTPUT_FILE = os.environ.get("CRON_OUTPUT_FILE", "/tmp/verify_pending_external.json")
+
+
+def _now() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _get_url(meta: dict[str, Any]) -> Optional[str]:
+    url = meta.get("url")
+    if isinstance(url, str) and url.startswith(("http://", "https://")):
+        return url
+    grounding = meta.get("grounding")
+    if isinstance(grounding, dict):
+        g_url = grounding.get("url")
+        if isinstance(g_url, str) and g_url.startswith(("http://", "https://")):
+            return g_url
+    return None
+
+
+def _http_status(url: str, timeout_s: float = 6.0) -> Optional[int]:
+    try:
+        import urllib.request
+
+        req = urllib.request.Request(
+            url,
+            method="GET",
+            headers={"User-Agent": "openclaw-secondbrain/verify_pending_external"},
+        )
+        with urllib.request.urlopen(req, timeout=timeout_s) as resp:
+            return int(getattr(resp, "status", 200))
+    except Exception:
+        return None
+
+
+def main() -> int:
+    if not DB_PATH.exists():
+        out = {"success": False, "error": f"db missing: {DB_PATH}", "time": _now()}
+        Path(OUTPUT_FILE).write_text(json.dumps(out, indent=2))
+        print(out["error"])
+        return 1
+
+    store = EngramStore(str(DB_PATH))
+    all_egs = []
+    offset = 0
+    while True:
+        batch = store.get_all(limit=2000, offset=offset)
+        if not batch:
+            break
+        all_egs.extend(batch)
+        offset += len(batch)
+    pending = [
+        eg
+        for eg in all_egs
+        if (not eg.correctness.is_final())
+    ]
+
+    confirmed = 0
+    rejected = 0
+    still_pending = 0
+    checked = 0
+
+    for eg in pending:
+        checked += 1
+        src = eg.metadata.get("source")
+        content = (eg.content or "").strip()
+
+        if src == "session" and (
+            content.startswith("Session Summary (sess_") or content.startswith("Please remember ")
+        ):
+            eg.correctness.reject("verify-pending", "Auto-reject: session placeholder")
+            store.save(eg)
+            rejected += 1
+            continue
+
+        if src == "web":
+            url = _get_url(eg.metadata)
+            if not url:
+                still_pending += 1
+                continue
+            status = _http_status(url)
+            if status is None:
+                still_pending += 1
+                continue
+            if 200 <= status < 300:
+                eg.correctness.confirm("verify-pending", f"Auto-confirm: web url ok ({status}) {url}")
+                store.save(eg)
+                confirmed += 1
+            else:
+                eg.correctness.reject("verify-pending", f"Auto-reject: web url status={status} {url}")
+                store.save(eg)
+                rejected += 1
+            continue
+
+        still_pending += 1
+
+    out = {
+        "success": True,
+        "time": _now(),
+        "total": len(all_egs),
+        "pending_before": len(pending),
+        "checked": checked,
+        "confirmed": confirmed,
+        "rejected": rejected,
+        "still_pending": still_pending,
+    }
+    Path(OUTPUT_FILE).write_text(json.dumps(out, indent=2))
+    print(
+        f"VERIFY: pending_before={out['pending_before']} confirmed={confirmed} rejected={rejected} still_pending={still_pending}"
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

docs/RELEASE_CHECKLIST.md

@@ -1,6 +1,6 @@
 # Second-Brain 2.0 — Release Checklist (Grundversion + FastAPI)
 
-Scope: OpenClaw workspace at `/root/.openclaw/workspace` with `second-brain/` and the shipped `systemd/` unit files.
+Scope: OpenClaw workspace at `/root/.openclaw/workspace` with `second-brain/` and the shipped `second-brain/systemd/` unit files.
 
 This checklist is copy/paste friendly. Run as a user with `sudo` (systemd commands need it).
 
@@ -24,13 +24,21 @@ ls -la /etc/systemd/system/openclaw-memory-archive.* 2>/dev/null || true
 If missing, install them (symlink is fine; copy is fine too):
 
 ```bash
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-secondbrain-*.service /etc/systemd/system/
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-secondbrain-*.timer /etc/systemd/system/
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-memory-archive.* /etc/systemd/system/
-sudo ln -sf /root/.openclaw/workspace/systemd/openclaw-secondbrain-dashboard.service /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-*.service /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-*.timer /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-memory-archive.* /etc/systemd/system/
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-dashboard.service /etc/systemd/system/
 sudo systemctl daemon-reload
 ```
 
+Optional (verification hardening):
+
+```bash
+sudo ln -sf /root/.openclaw/workspace/second-brain/systemd/openclaw-secondbrain-verify-pending.* /etc/systemd/system/
+sudo systemctl daemon-reload
+sudo systemctl enable --now openclaw-secondbrain-verify-pending.timer
+```
+
 ### Enable timers
 
 ```bash

docs/dashboard-ui-ux-design.md

@@ -0,0 +1,215 @@
+# Second-Brain Dashboard UI/UX Design Plan (mobile-first)
+
+Basis: `second-brain/templates/dashboard.html`, `second-brain/static/style.css` (aktueller Stand) + neue Graph-Controls aus `/tmp/second-brain-staging/`.
+
+## 1) Konkreter Design-Plan
+
+### Ziele (UX)
+- **Schnelles Triaging unterwegs:** Pending/Errors sofort sichtbar, 1-Hand-Bedienung.
+- **Konsistentes Design-System:** einheitliche Buttons/Inputs/Panels statt Einzellösungen.
+- **Graph als Diagnose-Tool:** klare Controls, Legende, nachvollziehbares Feedback (Loading/Empty/Errors).
+
+### Farbschema (Dark, high-contrast, "indigo + emerald")
+- **Background:** sehr dunkel (nahe #0f1117) für weniger Blendung.
+- **Surface (Cards/Panels):** abgestufte Flächen (Surface-1, Surface-2) für Hierarchie.
+- **Primary:** Indigo/Blue für interaktive Elemente und Highlights (bisher #6c8af5 bleibt als Basis).
+- **Success:** Emerald für Confirm/OK (bisher #3a7d3a → etwas heller/satter).
+- **Danger:** Red für Errors/Reject.
+- **Warning:** Amber für Pending/Match/Attention.
+- **Text:** fast-weiß, sekundär gedimmt.
+
+### Typografie
+- Systemfont-Stack (schnell, gut lesbar): `ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Arial`.
+- Skala (mobile-first):
+  - `text-xs` 12px (Labels, Meta)
+  - `text-sm` 13–14px (secondary)
+  - `text-md` 15–16px (Body)
+  - `text-lg` 18–20px (Titles)
+- **Ziffern:** optional `font-variant-numeric: tabular-nums;` für Stats.
+
+### Spacing/Rhythm
+- 4px Grid; Standard-Gaps: 8/12/16.
+- Container-Padding: 12px (mobile), 16–20px (>=768px).
+- Border-radius: 12–16px (Cards/Modals), 10px (Inputs/Buttons), 999px (Pills).
+
+### Komponenten (konkret)
+
+#### A) Tabs
+- Tabs bleiben als 3 Buttons, aber:
+  - **Active State** deutlicher (background + border + subtle glow).
+  - **Touch target** min. 44px Höhe.
+  - Sticky bleibt, aber mit leichter **Backdrop-Blurring** (wenn möglich) oder solid surface.
+
+#### B) Search
+- Search-Row wird als **kompakte Toolbar** gestaltet:
+  - Input + Filter in einer Zeile.
+  - Optional Quick-Chips darunter: `All / Pending / Confirmed / Errors` (klickbar) als Alternative zum Select.
+- Clear-Button (×) im Input (per CSS `::-webkit-search-cancel-button` oder eigener Button) für Mobile.
+
+#### C) Cards
+- Karte als 3 Zonen: Header (badges/tags/date), Body (content), Footer (actions).
+- Status wird stärker codiert:
+  - left-border + kleine **Status-Pill** (OK/Pending/Error) mit eindeutiger Farbe.
+- Body: bessere Lesbarkeit via `line-height: 1.55` und max-height/clamp optional.
+
+#### D) Modal
+- Modal als **Bottom Sheet** auf Mobile (>=50vh) + klassisches Center-Modal auf Desktop.
+- Close-Button größer + Tap-Area.
+- Inhalt in Tabs/Sections (History/Meta/Content) optional später.
+
+#### E) Graph + Controls (aus Staging)
+- Controls als **Control Bar** oberhalb Canvas:
+  - Primary: Physics toggle.
+  - Secondary: Fit, Reset, Reload.
+  - Text labels kurz (z.B. `Physics` statt `Physics: off`, state als Badge).
+- Canvas passt sich an Viewport an:
+  - `width: min(100%, 560px)`; Height: `min(65vh, 560px)` (CSS statt fester HTML Attribute, wenn möglich).
+- Legende als einklappbares Panel (`Details`/`summary`) oder leichtes Panel unter Canvas.
+
+#### F) Status Panels
+- Status-View nutzt vorhandene `.panel`/`.kv-*`:
+  - Gruppen: System, Storage, Jobs, Insights, Pending Queue.
+  - Jede Gruppe als Panel mit klarer Title Row.
+- Kritische Werte (Errors/Pending/Queue) farblich markieren.
+
+## 2) CSS-Variablen (Theme-Tokens) + Mapping
+
+### Token-Vorschlag (`:root`)
+```css
+:root {
+  /* typography */
+  --font-sans: ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Arial, sans-serif;
+
+  /* colors */
+  --bg: #0f1117;
+  --surface-1: #14151d;
+  --surface-2: #1a1b26;
+  --border: #2a2d3a;
+
+  --text: #e8e8ee;
+  --text-muted: #8b90a3;
+  --text-dim: #5c6276;
+
+  --primary: #6c8af5;      /* existing */
+  --primary-2: #8aa1ff;
+  --success: #2fbf71;
+  --danger: #ef4444;
+  --warning: #f7d154;      /* existing-ish */
+
+  /* shadows */
+  --shadow-1: 0 1px 0 rgba(0,0,0,.25), 0 8px 24px rgba(0,0,0,.35);
+
+  /* radius */
+  --r-sm: 10px;
+  --r-md: 14px;
+  --r-lg: 16px;
+
+  /* spacing */
+  --s-1: 4px;
+  --s-2: 8px;
+  --s-3: 12px;
+  --s-4: 16px;
+  --s-5: 20px;
+
+  /* control sizes */
+  --tap: 44px;
+}
+```
+
+### Mapping auf existierende Selektoren
+- `body`/`.app`
+  - `background: var(--bg)` statt `#141419`
+  - `font-family: var(--font-sans)`
+  - `color: var(--text)`
+- `.stats-bar`, `.tabs-bar`, `.search-box`
+  - `background: var(--surface-1)` (Stats ggf. Gradient bleibt, aber auf Tokens)
+  - `border-bottom: 1px solid var(--border)`
+- `.panel`, `.card`, `.modal-content`, `.graph-legend`
+  - `background: var(--surface-2)`
+  - `border: 1px solid var(--border)`
+  - `border-radius: var(--r-md)`
+- `.stat-num`, `#pageNum`, `.tab-btn.active`, `#searchInput:focus`
+  - `color/border-color: var(--primary)`
+- `.muted`, `.stat-label`, `.kv-key`, `.date`
+  - `color: var(--text-muted)` bzw. `var(--text-dim)`
+- Buttons
+  - vereinheitlichen über `.btn` + Modifier: `.btn.primary`, `.btn.danger`, `.btn.success`, `.btn.ghost`
+  - `min-height: var(--tap)` für Touch
+
+## 3) 5–10 priorisierte UI-Änderungen (mit Begründung)
+
+1) **Design Tokens (CSS vars) einführen** → reduziert Farbmix, erleichtert spätere Themes/Anpassungen.
+2) **Einheitliche Button-Komponente (`.btn`)** (inkl. `:active`, `:disabled`, `min-height`) → bessere Touch-UX, konsistente Interaktion.
+3) **Graph-Controls + Legende aus Staging in den Main-Template-Stand ziehen** → Graph wird tatsächlich bedienbar/selbsterklärend.
+4) **Responsive Graph-Canvas (CSS gesteuert)** statt fixer `width/height` → bessere Nutzung auf Phones, weniger Scroll.
+5) **Search als Toolbar + Clear-Action** → schnelleres Filtern unterwegs, weniger Friktion.
+6) **Modal als Bottom-Sheet auf Mobile** → angenehmer für längeren Content + History, weniger „winziges Fenster“.
+7) **Status/Health Werte farblich akzentuieren** (pending/errors/warn) → schnelleres Erkennen von Problemen.
+8) **Cards: Status-Pill + typografische Lesbarkeit** (line-height, spacing) → weniger „Textblock“, bessere Scanbarkeit.
+9) **Accessibility-Basics**: Focus-Rings, Kontrast, `prefers-reduced-motion` → weniger „invisible focus“ und bessere Bedienbarkeit.
+10) **Top-level Layout Max-Width für Desktop** (z.B. 560–720px) → verhindert „zu breite“ Zeilen.
+
+## 4) Optionale Patch-Vorschläge (Diff-Snippets, NICHT anwenden)
+
+> Hinweis: Snippets sind bewusst klein gehalten. Gesamt < 120 Zeilen.
+
+### Snippet A — Tokens + Button-System (style.css)
+```diff
+--- a/second-brain/static/style.css
++++ b/second-brain/static/style.css
+@@
++:root {
++  --font-sans: ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Arial, sans-serif;
++  --bg:#0f1117; --surface-1:#14151d; --surface-2:#1a1b26; --border:#2a2d3a;
++  --text:#e8e8ee; --text-muted:#8b90a3; --text-dim:#5c6276;
++  --primary:#6c8af5; --success:#2fbf71; --danger:#ef4444; --warning:#f7d154;
++  --r-sm:10px; --r-md:14px; --r-lg:16px;
++  --s-2:8px; --s-3:12px; --s-4:16px;
++  --tap:44px;
++}
++
++body { font-family: var(--font-sans); background: var(--bg); color: var(--text); }
++
++.btn{
++  min-height: var(--tap);
++  background: #1e1e28;
++  border: 1px solid var(--border);
++  border-radius: var(--r-sm);
++  padding: 8px 12px;
++  color: #cfd3ff;
++  font-weight: 700;
++}
++.btn.primary{ border-color: var(--primary); box-shadow: 0 0 0 1px rgba(108,138,245,0.18) inset; }
++.btn.success{ background: rgba(47,191,113,.18); border-color: rgba(47,191,113,.35); }
++.btn.danger{ background: rgba(239,68,68,.16); border-color: rgba(239,68,68,.35); }
++.btn:active{ transform: scale(.98); }
++.btn:disabled{ opacity: .45; }
+```
+
+### Snippet B — Graph Controls + Legend übernehmen (dashboard.html)
+```diff
+--- a/second-brain/templates/dashboard.html
++++ b/second-brain/templates/dashboard.html
+@@
+-        <div class="graph" id="graph" style="display:none;">
+-            <canvas id="graphCanvas" width="440" height="520"></canvas>
+-            <div class="muted small" id="graphHint">Lade Graph…</div>
+-        </div>
++        <div class="graph" id="graph" style="display:none;">
++            <div class="graph-controls">
++                <button class="btn primary" id="btnGraphPhysics" onclick="toggleGraphPhysics()">Physics</button>
++                <button class="btn" onclick="resetGraphView()">Reset</button>
++                <button class="btn" onclick="fitGraphView()">Fit</button>
++                <button class="btn" onclick="reloadGraph()">Reload</button>
++            </div>
++            <canvas id="graphCanvas" width="440" height="520"></canvas>
++            <div class="graph-hint muted small" id="graphHint">Lade Graph…</div>
++            <div class="graph-legend">
++                <div><strong>Graph</strong>: Zoom (Wheel/Pinch), Pan (Drag). Klick auf Engram öffnet Details, Klick auf Tag setzt Suche.</div>
++                <div class="legend-row"><span class="legend-dot engram"></span> Engram</div>
++                <div class="legend-row"><span class="legend-dot tag"></span> Tag</div>
++                <div class="legend-row"><span class="legend-dot match"></span> Match</div>
++            </div>
++        </div>
+```
+

fastapi_app.py

@@ -11,11 +11,14 @@ Goals:
 import json
 import os
 import sqlite3
+import subprocess
 from datetime import datetime, timezone
 from pathlib import Path
+from uuid import uuid4
+from urllib.parse import urlparse
 
 from fastapi import FastAPI, Form, Query, Request
-from fastapi.responses import HTMLResponse, JSONResponse, PlainTextResponse
+from fastapi.responses import HTMLResponse, JSONResponse, PlainTextResponse, StreamingResponse
 from fastapi.staticfiles import StaticFiles
 
 # ─── Config ──────────────────────────────────────────────────────────────────
@@ -51,11 +54,22 @@ def get_db():
 def parse_engram(row: sqlite3.Row) -> dict:
     meta = json.loads(row["metadata_json"] or "{}")
     correctness = json.loads(row["correctness_json"] or "{}")
+    verdict = correctness.get("verdict")
+    if not isinstance(verdict, str) or not verdict:
+        # Back-compat inference for older rows
+        if correctness.get("confirmed", False):
+            verdict = "confirmed_true"
+        elif int(correctness.get("rejections", 0) or 0) > 0:
+            verdict = "confirmed_false"
+        else:
+            verdict = "unknown"
     return {
         "id": row["id"],
         "content": row["content"],
         "confidence": meta.get("confidence", 0.0),
         "confirmed": correctness.get("confirmed", False),
+        "verdict": verdict,
+        "evidence": correctness.get("evidence", []),
         "confirmations": correctness.get("confirmations", 0),
         "rejections": correctness.get("rejections", 0),
         "tags": meta.get("tags", []),
@@ -69,6 +83,158 @@ def parse_engram(row: sqlite3.Row) -> dict:
     }
 
 
+def _now_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _update_correctness(engram_id: str, *, action: str, reason: str | None = None) -> dict:
+    """
+    Update correctness_json for an engram. action: confirm|reject
+    """
+    conn = get_db()
+    c = conn.cursor()
+    row = c.execute("SELECT correctness_json FROM engrams WHERE id = ?", (engram_id,)).fetchone()
+    if not row:
+        conn.close()
+        raise FileNotFoundError(f"Engram not found: {engram_id}")
+
+    corr = json.loads(row["correctness_json"] or "{}")
+    corr.setdefault("verdict", None)
+    corr.setdefault("evidence", [])
+    corr.setdefault("confirmed", False)
+    corr.setdefault("confirmations", 0)
+    corr.setdefault("rejections", 0)
+    corr.setdefault("review_history", [])
+    corr["last_reviewed"] = _now_iso()
+
+    entry = {
+        "by": "dashboard",
+        "action": action,
+        "at": corr["last_reviewed"],
+        "note": (reason or "").strip(),
+    }
+    try:
+        corr["review_history"].append(entry)
+    except Exception:
+        corr["review_history"] = [entry]
+
+    if action == "confirm":
+        corr["verdict"] = "confirmed_true"
+        corr["confirmed"] = True
+        corr["confirmations"] = int(corr.get("confirmations", 0) or 0) + 1
+    elif action == "reject":
+        corr["verdict"] = "confirmed_false"
+        corr["rejections"] = int(corr.get("rejections", 0) or 0) + 1
+        corr["confirmed"] = False
+
+    # Store minimal evidence for dashboard-driven actions.
+    try:
+        ev = corr.get("evidence")
+        if not isinstance(ev, list):
+            ev = []
+        ev.append(
+            {
+                "kind": "human",
+                "by": "dashboard",
+                "at": corr["last_reviewed"],
+                "action": action,
+            }
+        )
+        corr["evidence"] = ev[-50:]  # cap growth
+    except Exception:
+        pass
+
+    c.execute(
+        "UPDATE engrams SET correctness_json = ?, modified_at = ? WHERE id = ?",
+        (json.dumps(corr, ensure_ascii=False), corr["last_reviewed"], engram_id),
+    )
+    conn.commit()
+    conn.close()
+    return {"ok": True}
+
+
+def _bump_access(engram_id: str) -> dict:
+    conn = get_db()
+    c = conn.cursor()
+    row = c.execute("SELECT metadata_json FROM engrams WHERE id = ?", (engram_id,)).fetchone()
+    if not row:
+        conn.close()
+        raise FileNotFoundError(f"Engram not found: {engram_id}")
+    meta = json.loads(row["metadata_json"] or "{}")
+    meta["access_count"] = int(meta.get("access_count", 0) or 0) + 1
+    meta["last_accessed"] = _now_iso()
+    c.execute(
+        "UPDATE engrams SET metadata_json = ?, modified_at = ? WHERE id = ?",
+        (json.dumps(meta, ensure_ascii=False), meta["last_accessed"], engram_id),
+    )
+    conn.commit()
+    conn.close()
+    return {"ok": True}
+
+def _safe_json_extract_tags(meta_json: str) -> list[str]:
+    try:
+        d = json.loads(meta_json or "{}")
+        tags = d.get("tags") or []
+        return [t for t in tags if isinstance(t, str)]
+    except Exception:
+        return []
+
+def _host_from_meta(meta_json: str) -> str | None:
+    try:
+        d = json.loads(meta_json or "{}")
+        grounding = d.get("grounding")
+        url = d.get("url")
+        if isinstance(grounding, dict) and isinstance(grounding.get("url"), str):
+            url = grounding.get("url")
+        if not isinstance(url, str):
+            return None
+        parsed = urlparse(url)
+        return parsed.hostname
+    except Exception:
+        return None
+
+def _systemd_unit_state(unit: str) -> dict:
+    """
+    Best-effort systemd status snapshot for a known unit.
+    Never raises; returns minimal fields.
+    """
+    try:
+        out = subprocess.check_output(
+            ["systemctl", "show", unit, "--no-page", "--property=ActiveState,SubState,Result,ExecMainStatus,ExecMainCode,ExecMainStartTimestamp,ExecMainExitTimestamp"],
+            text=True,
+            stderr=subprocess.STDOUT,
+            timeout=2,
+        )
+        kv = {}
+        for line in out.splitlines():
+            if "=" in line:
+                k, v = line.split("=", 1)
+                kv[k] = v
+        return {
+            "unit": unit,
+            "active": kv.get("ActiveState"),
+            "sub": kv.get("SubState"),
+            "result": kv.get("Result"),
+            "exit_status": kv.get("ExecMainStatus"),
+            "start_ts": kv.get("ExecMainStartTimestamp"),
+            "exit_ts": kv.get("ExecMainExitTimestamp"),
+        }
+    except Exception as e:
+        return {"unit": unit, "error": str(e)}
+
+def _dir_size_bytes(path: Path) -> int:
+    total = 0
+    try:
+        for p in path.rglob("*"):
+            try:
+                if p.is_file():
+                    total += p.stat().st_size
+            except Exception:
+                pass
+    except Exception:
+        pass
+    return total
+
 # ─── API Endpoints ───────────────────────────────────────────────────────────
 
 @app.get("/healthz", response_class=PlainTextResponse)
@@ -83,6 +249,363 @@ def api_config():
         "db_path": str(DB_PATH),
     }
 
+@app.get("/api/db_info")
+def api_db_info():
+    if not DB_PATH.exists():
+        raise FileNotFoundError(f"DB not found: {DB_PATH}")
+    st = DB_PATH.stat()
+    return {
+        "db_path": str(DB_PATH),
+        "size_bytes": st.st_size,
+        "mtime": datetime.fromtimestamp(st.st_mtime, tz=timezone.utc).isoformat(),
+    }
+
+@app.get("/api/storage_stats")
+def api_storage_stats():
+    conn = get_db()
+    c = conn.cursor()
+    total = c.execute("SELECT COUNT(*) FROM engrams").fetchone()[0]
+    confirmed_true = c.execute(
+        """
+        SELECT COUNT(*) FROM engrams
+        WHERE (
+          json_extract(correctness_json, '$.verdict') = 'confirmed_true'
+          OR (json_extract(correctness_json, '$.verdict') IS NULL AND json_extract(correctness_json, '$.confirmed') = 1)
+        )
+        """
+    ).fetchone()[0]
+    confirmed_false = c.execute(
+        """
+        SELECT COUNT(*) FROM engrams
+        WHERE (
+          json_extract(correctness_json, '$.verdict') = 'confirmed_false'
+          OR (json_extract(correctness_json, '$.verdict') IS NULL
+              AND json_extract(correctness_json, '$.confirmed') = 0
+              AND COALESCE(json_extract(correctness_json, '$.rejections'), 0) > 0)
+        )
+        """
+    ).fetchone()[0]
+    sources = {
+        r[0]: r[1]
+        for r in c.execute(
+            "SELECT json_extract(metadata_json, '$.source') AS src, COUNT(*) FROM engrams GROUP BY src ORDER BY COUNT(*) DESC"
+        ).fetchall()
+        if r[0] is not None
+    }
+    conn.close()
+
+    chroma_dir = WORKSPACE / "data" / "chroma"
+    emb_cache_dir = WORKSPACE / "data" / "embedding_cache"
+    vec_state_path = WORKSPACE / "data" / "vector_index_state.json"
+
+    vec_state = {}
+    if vec_state_path.exists():
+        try:
+            vec_state = json.loads(vec_state_path.read_text())
+        except Exception:
+            vec_state = {}
+
+    obsidian_cfg_path = WORKSPACE / "data" / "obsidian_config.json"
+    obsidian_cfg = None
+    if obsidian_cfg_path.exists():
+        try:
+            obsidian_cfg = json.loads(obsidian_cfg_path.read_text())
+        except Exception:
+            obsidian_cfg = {"raw": obsidian_cfg_path.read_text()[:2000]}
+
+    backup_files = sorted((WORKSPACE / "data").glob("backup_*.jsonl"))
+
+    return {
+        "sql": {
+            "total_engrams": total,
+            "confirmed": confirmed_true,
+            "rejected": confirmed_false,
+            "pending": total - confirmed_true - confirmed_false,
+            "by_source": sources,
+        },
+        "vector": {
+            "chroma_dir": str(chroma_dir),
+            "chroma_size_bytes": _dir_size_bytes(chroma_dir) if chroma_dir.exists() else 0,
+            "embedding_cache_dir": str(emb_cache_dir),
+            "embedding_cache_files": len(list(emb_cache_dir.glob("*.json"))) if emb_cache_dir.exists() else 0,
+            "vector_state": vec_state,
+        },
+        "obsidian": {
+            "config_path": str(obsidian_cfg_path),
+            "configured": bool(obsidian_cfg),
+            "config": obsidian_cfg,
+        },
+        "backups": {
+            "count": len(backup_files),
+            "latest": str(backup_files[-1]) if backup_files else None,
+        },
+    }
+
+@app.get("/api/jobs")
+def api_jobs():
+    # Known units that influence "freshness" of the brain.
+    units = [
+        "openclaw-secondbrain-ingest-memory.service",
+        "openclaw-secondbrain-index-vectors.service",
+        "openclaw-secondbrain-review.service",
+        "openclaw-secondbrain-heartbeat.service",
+        "openclaw-secondbrain-verify-pending.service",
+    ]
+    return {"items": [_systemd_unit_state(u) for u in units]}
+
+@app.get("/api/insights")
+def api_insights(limit: int = Query(8, ge=1, le=50)):
+    conn = get_db()
+    c = conn.cursor()
+    rows = c.execute(
+        "SELECT id, metadata_json, correctness_json, created_at, modified_at FROM engrams ORDER BY created_at DESC LIMIT 2000"
+    ).fetchall()
+    total = c.execute("SELECT COUNT(*) FROM engrams").fetchone()[0]
+    confirmed_true = c.execute(
+        """
+        SELECT COUNT(*) FROM engrams
+        WHERE (
+          json_extract(correctness_json, '$.verdict') = 'confirmed_true'
+          OR (json_extract(correctness_json, '$.verdict') IS NULL AND json_extract(correctness_json, '$.confirmed') = 1)
+        )
+        """
+    ).fetchone()[0]
+    confirmed_false = c.execute(
+        """
+        SELECT COUNT(*) FROM engrams
+        WHERE (
+          json_extract(correctness_json, '$.verdict') = 'confirmed_false'
+          OR (json_extract(correctness_json, '$.verdict') IS NULL
+              AND json_extract(correctness_json, '$.confirmed') = 0
+              AND COALESCE(json_extract(correctness_json, '$.rejections'), 0) > 0)
+        )
+        """
+    ).fetchone()[0]
+    pending = total - confirmed_true - confirmed_false
+
+    tag_counts: dict[str, int] = {}
+    source_counts: dict[str, int] = {}
+    host_counts: dict[str, int] = {}
+    active: list[dict] = []
+    forgotten: list[dict] = []
+
+    for r in rows:
+        try:
+            meta = json.loads(r["metadata_json"] or "{}")
+        except Exception:
+            meta = {}
+        src = meta.get("source", "unknown")
+        source_counts[src] = source_counts.get(src, 0) + 1
+        for t in (meta.get("tags") or []):
+            if isinstance(t, str):
+                tag_counts[t] = tag_counts.get(t, 0) + 1
+        try:
+            host = _host_from_meta(r["metadata_json"])
+        except Exception:
+            host = None
+        if host:
+            host_counts[host] = host_counts.get(host, 0) + 1
+
+        access_count = int(meta.get("access_count", 0) or 0)
+        created = meta.get("created", r["created_at"])
+        if access_count >= 5 and len(active) < limit:
+            active.append(
+                {
+                    "id": r["id"],
+                    "access_count": access_count,
+                    "source": src,
+                    "created": created,
+                }
+            )
+        if access_count == 0 and len(forgotten) < limit:
+            forgotten.append(
+                {
+                    "id": r["id"],
+                    "access_count": access_count,
+                    "source": src,
+                    "created": created,
+                }
+            )
+
+    def top_k(d: dict[str, int]) -> list[dict]:
+        return [
+            {"key": k, "count": v}
+            for k, v in sorted(d.items(), key=lambda kv: kv[1], reverse=True)[:limit]
+        ]
+
+    conn.close()
+    return {
+        "total": total,
+        "confirmed": confirmed_true,
+        "rejected": confirmed_false,
+        "pending": pending,
+        "top_tags": top_k(tag_counts),
+        "top_sources": top_k(source_counts),
+        "top_hosts": top_k(host_counts),
+        "active_engrams": active,
+        "forgotten_engrams": forgotten,
+    }
+
+@app.get("/api/graph")
+def api_graph(
+    limit_nodes: int = Query(0, ge=0, le=50000),
+    limit_edges: int = Query(0, ge=0, le=200000),
+):
+    """
+    Returns a lightweight graph view:
+    - Nodes: engrams + tag:<tag> + host:<hostname>
+    - Edges: engram->tag and engram->host plus explicit engrams_links edges.
+    """
+    conn = get_db()
+    c = conn.cursor()
+    if limit_nodes > 0:
+        # Fetch a bigger window than the final node cap so trim can keep hubs + neighbors.
+        engram_fetch = min(50000, max(1000, int(limit_nodes * 3)))
+    else:
+        engram_fetch = None
+
+    if limit_edges > 0:
+        link_fetch = limit_edges
+    elif limit_nodes > 0:
+        link_fetch = min(200000, max(2000, int(limit_nodes * 10)))
+    else:
+        link_fetch = None
+
+    if engram_fetch is None:
+        rows = c.execute("SELECT id, metadata_json, correctness_json, created_at, modified_at FROM engrams ORDER BY created_at DESC").fetchall()
+    else:
+        rows = c.execute(
+            "SELECT id, metadata_json, correctness_json, created_at, modified_at FROM engrams ORDER BY created_at DESC LIMIT ?",
+            (engram_fetch,),
+        ).fetchall()
+
+    if link_fetch is None:
+        link_rows = c.execute("SELECT from_id, to_id FROM engrams_links ORDER BY rowid DESC").fetchall()
+    else:
+        link_rows = c.execute(
+            "SELECT from_id, to_id FROM engrams_links ORDER BY rowid DESC LIMIT ?",
+            (link_fetch,),
+        ).fetchall()
+    conn.close()
+
+    nodes: dict[str, dict] = {}
+    edges: list[dict] = []
+
+    def add_node(nid: str, kind: str, label: str | None = None, weight: float | None = None):
+        if nid not in nodes:
+            nodes[nid] = {"id": nid, "kind": kind}
+        if label is not None:
+            nodes[nid]["label"] = label
+        if weight is not None:
+            nodes[nid]["weight"] = weight
+
+    for r in rows:
+        eid = r["id"]
+        try:
+            meta = json.loads(r["metadata_json"] or "{}")
+        except Exception:
+            meta = {}
+        try:
+            corr = json.loads(r["correctness_json"] or "{}")
+        except Exception:
+            corr = {}
+        verdict = corr.get("verdict")
+        if not isinstance(verdict, str) or not verdict:
+            if corr.get("confirmed", False):
+                verdict = "confirmed_true"
+            elif int(corr.get("rejections", 0) or 0) > 0:
+                verdict = "confirmed_false"
+            else:
+                verdict = "unknown"
+
+        add_node(
+            eid,
+            "engram",
+            label=eid[:8],
+            weight=float(meta.get("access_count", 0) or 0),
+        )
+        nodes[eid].update(
+            {
+                "source": meta.get("source", "unknown"),
+                "confidence": float(meta.get("confidence", 0.0) or 0.0),
+                "created": meta.get("created", r["created_at"]),
+                "modified": meta.get("modified", r["modified_at"]),
+                "last_accessed": meta.get("last_accessed"),
+                "verdict": verdict,
+                "confirmed": bool(corr.get("confirmed", False)),
+                "rejections": int(corr.get("rejections", 0) or 0),
+                "grounding": meta.get("grounding", 0),
+                "predict_locked": bool(meta.get("predict_locked", False)),
+            }
+        )
+        for t in _safe_json_extract_tags(r["metadata_json"]):
+            tid = f"tag:{t}"
+            add_node(tid, "tag", label=t)
+            edges.append({"from": eid, "to": tid, "kind": "has_tag", "weight": 0.35})
+        host = _host_from_meta(r["metadata_json"])
+        if host:
+            hid = f"host:{host}"
+            add_node(hid, "host", label=host)
+            edges.append({"from": eid, "to": hid, "kind": "grounded_at", "weight": 0.25})
+
+    for fr, to in link_rows:
+        add_node(fr, "engram")
+        add_node(to, "engram")
+        edges.append({"from": fr, "to": to, "kind": "link", "weight": 1.0})
+
+    # Trim nodes to keep payload bounded (prefer engrams and connected tags/hosts)
+    if limit_nodes > 0 and len(nodes) > limit_nodes:
+        # Keep a balanced subset: many engrams plus the most-connected non-engrams.
+        kept: dict[str, dict] = {}
+        engram_budget = int(limit_nodes * 0.7)
+
+        # 1) Keep newest engrams first (they appear first in `rows` loop insertion order)
+        for r in rows:
+            eid = r["id"]
+            if eid in nodes:
+                kept[eid] = nodes[eid]
+            if len(kept) >= engram_budget:
+                break
+
+        # 2) Rank remaining nodes by degree within current edge set
+        degree: dict[str, int] = {}
+        for e in edges:
+            degree[e["from"]] = degree.get(e["from"], 0) + 1
+            degree[e["to"]] = degree.get(e["to"], 0) + 1
+
+        remaining = [nid for nid in nodes.keys() if nid not in kept]
+        remaining.sort(key=lambda nid: degree.get(nid, 0), reverse=True)
+        for nid in remaining:
+            kept[nid] = nodes[nid]
+            if len(kept) >= limit_nodes:
+                break
+
+        nodes = kept
+        edges = [e for e in edges if e["from"] in nodes and e["to"] in nodes]
+
+    return {"nodes": list(nodes.values()), "edges": edges}
+
+@app.get("/api/events")
+def api_events():
+    """
+    Server-Sent Events stream for lightweight real-time UI refresh.
+    """
+    import time
+
+    def gen():
+        while True:
+            payload = {
+                "ts": datetime.now(timezone.utc).isoformat(),
+                "stats": api_stats(),
+                "storage": api_storage_stats(),
+                "jobs": api_jobs(),
+                "insights": api_insights(limit=8),
+            }
+            yield f"data: {json.dumps(payload, ensure_ascii=False)}\n\n"
+            time.sleep(5)
+
+    return StreamingResponse(gen(), media_type="text/event-stream")
+
 
 @app.exception_handler(FileNotFoundError)
 def handle_file_not_found(request: Request, exc: FileNotFoundError):
@@ -105,10 +628,27 @@ def api_stats():
     conn = get_db()
     c = conn.cursor()
     total = c.execute("SELECT COUNT(*) FROM engrams").fetchone()[0]
-    confirmed = c.execute(
-        "SELECT COUNT(*) FROM engrams WHERE json_extract(correctness_json, '$.confirmed') = 1"
+    confirmed_true = c.execute(
+        """
+        SELECT COUNT(*) FROM engrams
+        WHERE (
+          json_extract(correctness_json, '$.verdict') = 'confirmed_true'
+          OR (json_extract(correctness_json, '$.verdict') IS NULL AND json_extract(correctness_json, '$.confirmed') = 1)
+        )
+        """
     ).fetchone()[0]
-    pending = total - confirmed
+    confirmed_false = c.execute(
+        """
+        SELECT COUNT(*) FROM engrams
+        WHERE (
+          json_extract(correctness_json, '$.verdict') = 'confirmed_false'
+          OR (json_extract(correctness_json, '$.verdict') IS NULL
+              AND json_extract(correctness_json, '$.confirmed') = 0
+              AND COALESCE(json_extract(correctness_json, '$.rejections'), 0) > 0)
+        )
+        """
+    ).fetchone()[0]
+    pending = total - confirmed_true - confirmed_false
     errors = c.execute(
         "SELECT COUNT(*) FROM engrams WHERE json_extract(metadata_json, '$.tags') LIKE '%error%'"
     ).fetchone()[0]
@@ -118,7 +658,8 @@ def api_stats():
     conn.close()
     return {
         "total": total,
-        "confirmed": confirmed,
+        "confirmed": confirmed_true,
+        "rejected": confirmed_false,
         "pending": pending,
         "errors": errors,
         "avg_confidence": round(avg_conf, 2),
@@ -131,6 +672,7 @@ def api_engrams(
     offset: int = Query(0, ge=0),
     tag: str = Query(None),
     confirmed: bool = Query(None),
+    verdict: str = Query(None),
     search: str = Query(None),
     min_confidence: float = Query(0.0),
 ):
@@ -144,9 +686,30 @@ def api_engrams(
         params.append(f'%"{tag}"%')
 
     if confirmed is not None:
+        if confirmed:
+            # confirmed == statement is true (verdict confirmed_true)
             where_clauses.append(
-            f"json_extract(correctness_json, '$.confirmed') = {int(confirmed)}"
+                "("
+                "json_extract(correctness_json, '$.verdict') = 'confirmed_true' "
+                "OR (json_extract(correctness_json, '$.verdict') IS NULL AND json_extract(correctness_json, '$.confirmed') = 1)"
+                ")"
             )
+        else:
+            # pending/unresolved (unknown/probable) but exclude confirmed_false.
+            where_clauses.append(
+                "("
+                "json_extract(correctness_json, '$.verdict') IN ('unknown','probable_true','probable_false') "
+                "OR (json_extract(correctness_json, '$.verdict') IS NULL "
+                "    AND json_extract(correctness_json, '$.confirmed') = 0 "
+                "    AND COALESCE(json_extract(correctness_json, '$.rejections'), 0) = 0)"
+                ")"
+            )
+
+    if verdict:
+        v = verdict.strip()
+        if v in ("unknown", "probable_true", "probable_false", "confirmed_true", "confirmed_false"):
+            where_clauses.append("json_extract(correctness_json, '$.verdict') = ?")
+            params.append(v)
 
     if search:
         # Use FTS
@@ -204,70 +767,107 @@ def api_engram_detail(engram_id: str):
     return result
 
 
-@app.post("/api/engrams/{engram_id}/confirm")
-def api_confirm(engram_id: str, reason: str = Form("")):
+@app.get("/api/pending")
+def api_pending(
+    limit: int = Query(20, ge=1, le=200),
+    offset: int = Query(0, ge=0),
+    source: str | None = Query(None),
+):
     conn = get_db()
     c = conn.cursor()
-    row = c.execute(
-        "SELECT correctness_json FROM engrams WHERE id = ?", (engram_id,)
-    ).fetchone()
-    if not row:
+    where = ["json_extract(correctness_json, '$.confirmed') = 0"]
+    params: list = []
+    if source:
+        where.append("json_extract(metadata_json, '$.source') = ?")
+        params.append(source)
+
+    rows = c.execute(
+        f"""
+        SELECT * FROM engrams
+        WHERE {' AND '.join(where)}
+        ORDER BY created_at DESC
+        LIMIT ? OFFSET ?
+        """,
+        params + [limit, offset],
+    ).fetchall()
+    items = [parse_engram(r) for r in rows]
     conn.close()
-        return JSONResponse({"error": "Not found"}, status_code=404)
+    return {"items": items, "limit": limit, "offset": offset}
 
-    correctness = json.loads(row["correctness_json"] or "{}")
-    correctness["confirmed"] = True
-    correctness["confirmations"] = correctness.get("confirmations", 0) + 1
-    correctness["last_reviewed"] = datetime.now(timezone.utc).isoformat()
-    review_history = correctness.get("review_history", [])
-    review_history.append({
-        "by": "web",
-        "action": "confirm",
-        "at": datetime.now(timezone.utc).isoformat(),
-        "note": reason or "confirmed via dashboard",
-    })
-    correctness["review_history"] = review_history
 
+@app.post("/api/engrams")
+def api_create_engram(content: str = Form(...), tags: str = Form("")):
+    content = (content or "").strip()
+    if not content:
+        return JSONResponse({"error": "content required"}, status_code=400)
+    tag_list = [t.strip() for t in (tags or "").split(",") if t.strip()]
+    now = _now_iso()
+    engram_id = str(uuid4())
+    meta = {
+        "source": "user",
+        "confidence": 0.7,
+        "created": now,
+        "modified": now,
+        "access_count": 0,
+        "last_accessed": now,
+        "tags": tag_list,
+        "session_id": None,
+        "agent_id": None,
+        "grounding": 0,
+    }
+    corr = {
+        "confirmed": False,
+        "confirmations": 0,
+        "rejections": 0,
+        "last_reviewed": None,
+        "review_history": [],
+    }
+    conn = get_db()
+    c = conn.cursor()
     c.execute(
-        "UPDATE engrams SET correctness_json = ?, modified_at = ? WHERE id = ?",
-        (json.dumps(correctness), datetime.now(timezone.utc).isoformat(), engram_id),
+        """
+        INSERT INTO engrams (id, content, metadata_json, correctness_json, links_json, hierarchy_json, embedding_json, created_at, modified_at)
+        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+        """,
+        (
+            engram_id,
+            content,
+            json.dumps(meta, ensure_ascii=False),
+            json.dumps(corr, ensure_ascii=False),
+            "[]",
+            "{}",
+            None,
+            now,
+            now,
+        ),
     )
     conn.commit()
     conn.close()
-    return {"success": True, "engram_id": engram_id}
+    return {"id": engram_id}
+
+
+@app.post("/api/engrams/{engram_id}/confirm")
+def api_confirm_engram(engram_id: str, reason: str = Form("")):
+    try:
+        return _update_correctness(engram_id, action="confirm", reason=reason)
+    except FileNotFoundError as e:
+        return JSONResponse({"error": str(e)}, status_code=404)
 
 
 @app.post("/api/engrams/{engram_id}/reject")
-def api_reject(engram_id: str, reason: str = Form("")):
-    conn = get_db()
-    c = conn.cursor()
-    row = c.execute(
-        "SELECT correctness_json FROM engrams WHERE id = ?", (engram_id,)
-    ).fetchone()
-    if not row:
-        conn.close()
-        return JSONResponse({"error": "Not found"}, status_code=404)
+def api_reject_engram(engram_id: str, reason: str = Form("")):
+    try:
+        return _update_correctness(engram_id, action="reject", reason=reason)
+    except FileNotFoundError as e:
+        return JSONResponse({"error": str(e)}, status_code=404)
 
-    correctness = json.loads(row["correctness_json"] or "{}")
-    correctness["confirmed"] = False
-    correctness["rejections"] = correctness.get("rejections", 0) + 1
-    correctness["last_reviewed"] = datetime.now(timezone.utc).isoformat()
-    review_history = correctness.get("review_history", [])
-    review_history.append({
-        "by": "web",
-        "action": "reject",
-        "at": datetime.now(timezone.utc).isoformat(),
-        "note": reason or "rejected via dashboard",
-    })
-    correctness["review_history"] = review_history
 
-    c.execute(
-        "UPDATE engrams SET correctness_json = ?, modified_at = ? WHERE id = ?",
-        (json.dumps(correctness), datetime.now(timezone.utc).isoformat(), engram_id),
-    )
-    conn.commit()
-    conn.close()
-    return {"success": True, "engram_id": engram_id}
+@app.post("/api/engrams/{engram_id}/refresh")
+def api_refresh_engram(engram_id: str):
+    try:
+        return _bump_access(engram_id)
+    except FileNotFoundError as e:
+        return JSONResponse({"error": str(e)}, status_code=404)
 
 
 @app.post("/api/engrams/{engram_id}/refresh")
@@ -320,6 +920,8 @@ def api_create_engram(content: str = Form(...), tags: str = Form(""), source: st
         "hash": "",
     }
     correctness = {
+        "verdict": "unknown",
+        "evidence": [],
         "confirmed": False,
         "confirmations": 0,
         "rejections": 0,
@@ -347,7 +949,12 @@ def api_pending(limit: int = Query(20, ge=1, le=100), offset: int = Query(0, ge=
     rows = c.execute(
         """
         SELECT * FROM engrams
-        WHERE json_extract(correctness_json, '$.confirmed') = 0
+        WHERE (
+          json_extract(correctness_json, '$.verdict') IN ('unknown','probable_true','probable_false')
+          OR (json_extract(correctness_json, '$.verdict') IS NULL
+              AND json_extract(correctness_json, '$.confirmed') = 0
+              AND COALESCE(json_extract(correctness_json, '$.rejections'), 0) = 0)
+        )
         ORDER BY created_at DESC
         LIMIT ? OFFSET ?
         """,

scripts/import_web_design_markers.py

@@ -0,0 +1,191 @@
+#!/usr/bin/env python3
+import argparse
+import json
+import hashlib
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional, Tuple
+
+from src.engram import Engram, Grounding
+from src.store import EngramStore
+
+
+def _now_utc_iso() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _hash16(text: str) -> str:
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()[:16]
+
+
+def _iter_jsonl(path: Path) -> Iterable[Dict[str, Any]]:
+    with path.open("r", encoding="utf-8") as f:
+        for line_no, line in enumerate(f, start=1):
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                obj = json.loads(line)
+            except Exception:
+                raise SystemExit(f"Invalid JSON at {path}:{line_no}")
+            if not isinstance(obj, dict):
+                continue
+            yield obj
+
+
+def _marker_to_content(marker_obj: Dict[str, Any]) -> Tuple[str, List[Dict[str, Any]]]:
+    marker = str(marker_obj.get("marker", "")).strip()
+    details = str(marker_obj.get("details", "")).strip()
+    checks = marker_obj.get("checks") or []
+    sources = marker_obj.get("sources") or []
+
+    if not marker:
+        raise ValueError("missing marker")
+
+    evidence: List[Dict[str, Any]] = []
+    for src in sources:
+        if not isinstance(src, dict):
+            continue
+        url = (src.get("url") or "").strip()
+        title = (src.get("title") or "").strip()
+        if not url:
+            continue
+        evidence.append({"url": url, "title": title})
+
+    lines: List[str] = []
+    lines.append(f"WEBDEV_MARKER: {marker}")
+    if details:
+        lines.append("")
+        lines.append(f"Details: {details}")
+    if isinstance(checks, list) and checks:
+        lines.append("")
+        lines.append("Checks:")
+        for c in checks[:8]:
+            c = str(c).strip()
+            if c:
+                lines.append(f"- {c}")
+    if evidence:
+        lines.append("")
+        lines.append("Sources:")
+        for ev in evidence[:12]:
+            title = (ev.get("title") or "").strip()
+            url = (ev.get("url") or "").strip()
+            if title:
+                lines.append(f"- {title}: {url}")
+            else:
+                lines.append(f"- {url}")
+    return "\n".join(lines).strip(), evidence
+
+
+def _tags_for(marker_obj: Dict[str, Any]) -> List[str]:
+    tags = ["web_design", "web_development", "mobile"]
+    area = str(marker_obj.get("area", "")).strip()
+    if area:
+        tags.append(area)
+    return tags
+
+
+def import_markers(
+    db_path: Path,
+    jsonl_paths: List[Path],
+    source: str,
+    verdict: str,
+    agent_id: str,
+    dry_run: bool,
+) -> Dict[str, int]:
+    store = EngramStore(str(db_path))
+
+    stats = {"seen": 0, "imported": 0, "skipped_dup": 0, "skipped_invalid": 0}
+    seen_hashes: set[str] = set()
+
+    # Preload existing hashes (fast-ish; avoids duplicate spam).
+    existing_hashes: set[str] = set()
+    try:
+        cur = store._conn.execute("SELECT metadata_json FROM engrams")  # noqa: SLF001
+        for row in cur.fetchall():
+            try:
+                meta = json.loads(row["metadata_json"])
+                h = meta.get("hash")
+                if isinstance(h, str) and h:
+                    existing_hashes.add(h)
+            except Exception:
+                continue
+    except Exception:
+        # If this fails (schema mismatch), proceed without preload.
+        existing_hashes = set()
+
+    for path in jsonl_paths:
+        for marker_obj in _iter_jsonl(path):
+            if (marker_obj.get("kind") or "") != "web_design_marker":
+                continue
+            stats["seen"] += 1
+            try:
+                content, evidence = _marker_to_content(marker_obj)
+            except Exception:
+                stats["skipped_invalid"] += 1
+                continue
+
+            h = _hash16(content)
+            if h in seen_hashes or h in existing_hashes:
+                stats["skipped_dup"] += 1
+                continue
+            seen_hashes.add(h)
+
+            eg = Engram.create(
+                content=content,
+                source=source,
+                confidence=0.75,
+                tags=_tags_for(marker_obj),
+                session_id=None,
+                agent_id=agent_id or str(marker_obj.get("agent_id") or ""),
+                grounding=Grounding.SOURCED,
+            )
+            # Overwrite hash to exactly match our content representation.
+            eg.metadata["hash"] = h
+            eg.metadata["modified"] = _now_utc_iso()
+            eg.metadata["created"] = marker_obj.get("created_at") or eg.metadata["created"]
+
+            eg.correctness.set_verdict(
+                by=agent_id or "importer",
+                verdict=verdict,
+                note=f"Imported from {path.name}",
+                evidence=evidence,
+            )
+
+            if not dry_run:
+                store.save(eg)
+            stats["imported"] += 1
+
+    return stats
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description="Import web_design_marker JSONL files into brain.sqlite")
+    p.add_argument("--db", default="second-brain/data/brain.sqlite", help="Path to brain.sqlite")
+    p.add_argument("--glob", default="/tmp/web_design_markers_*.jsonl", help="Glob for marker JSONL files")
+    p.add_argument("--source", default="web_research", help="Engram source")
+    p.add_argument("--verdict", default="probable_true", help="Correctness verdict")
+    p.add_argument("--agent-id", default="web_research_import", help="Agent id to record")
+    p.add_argument("--dry-run", action="store_true", help="Parse/dedupe but do not write to DB")
+    args = p.parse_args()
+
+    db_path = Path(args.db)
+    jsonl_paths = sorted(Path("/").glob(args.glob.lstrip("/"))) if args.glob.startswith("/") else sorted(Path(".").glob(args.glob))
+    if not jsonl_paths:
+        raise SystemExit(f"No files match glob: {args.glob}")
+
+    stats = import_markers(
+        db_path=db_path,
+        jsonl_paths=jsonl_paths,
+        source=args.source,
+        verdict=args.verdict,
+        agent_id=args.agent_id,
+        dry_run=bool(args.dry_run),
+    )
+
+    print(json.dumps({"db": str(db_path), "files": [str(p) for p in jsonl_paths], "stats": stats}, ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+    main()
+

scripts/process_pending_engrams.py

@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+"""
+Process pending second brain engrams.
+- For unconfirmed, unrejected engrams: evaluate confidence
+- If confidence > 0.8: confirm
+- If confidence < 0.3: reject
+- Otherwise: mark for review (leave as is)
+- Check for stale topics and archive if needed
+- Produce summary report
+"""
+
+import sys
+import json
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Add src to path and set PYTHONPATH for proper module resolution
+base_dir = Path(__file__).parent.parent
+sys.path.insert(0, str(base_dir / "src"))
+
+# Import using absolute module paths
+from src.store import EngramStore
+from src.engram import Engram, Grounding
+
+DB_PATH = Path(__file__).parent.parent / "data" / "brain.sqlite"
+
+
+def is_stale(engram: Engram, days_threshold: int = 90) -> bool:
+    """Check if an engram is stale (old and rarely accessed)."""
+    created = engram.metadata.get("created", "")
+    access_count = engram.metadata.get("access_count", 0)
+    last_accessed = engram.metadata.get("last_accessed", created)
+
+    try:
+        created_dt = datetime.fromisoformat(created.replace("Z", "+00:00"))
+        last_accessed_dt = datetime.fromisoformat(last_accessed.replace("Z", "+00:00"))
+        age_days = (datetime.now(timezone.utc) - created_dt).total_seconds() / 86400
+        days_since_access = (datetime.now(timezone.utc) - last_accessed_dt).total_seconds() / 86400
+
+        # Stale if: old (>90 days) AND rarely accessed (<3 times) AND not accessed recently (>60 days)
+        if age_days > days_threshold and access_count < 3 and days_since_access > 60:
+            return True
+    except Exception:
+        pass
+
+    return False
+
+
+def process_pending_engrams():
+    """Main processing function."""
+    store = EngramStore(str(DB_PATH))
+
+    # Get all engrams
+    all_engrams = store.get_all(limit=10000)
+    print(f"Total engrams in database: {len(all_engrams)}")
+
+    # Filter pending (unconfirmed and unrejected)
+    # Unconfirmed: not confirmed_true, not confirmed_false
+    pending = []
+    for eg in all_engrams:
+        verdict = eg.correctness.verdict
+        if verdict not in ("confirmed_true", "confirmed_false"):
+            pending.append(eg)
+
+    print(f"Pending engrams (unconfirmed/unrejected): {len(pending)}")
+
+    actions = {
+        "confirmed": 0,
+        "rejected": 0,
+        "left_for_review": 0,
+        "archived_stale": 0,
+        "errors": 0
+    }
+
+    details = []
+
+    for eg in pending:
+        try:
+            confidence = eg.compute_confidence()
+            engram_id = str(eg.id)
+            content_preview = eg.content[:80] + ("..." if len(eg.content) > 80 else "")
+
+            # Check if stale and should be archived
+            if is_stale(eg):
+                # For stale engrams, we'll mark them in metadata for archiving
+                # Instead of deleting, we'll add an "archived" tag and lower their priority
+                tags = eg.metadata.get("tags", [])
+                if "archived" not in tags:
+                    tags.append("archived")
+                    eg.metadata["tags"] = tags
+                    eg.metadata["archived_at"] = datetime.now(timezone.utc).isoformat()
+                    store.save(eg)
+                    actions["archived_stale"] += 1
+                    details.append(f"📦 Archived stale: [{engram_id[:8]}] {content_preview} (conf: {confidence:.2f})")
+                # Even if stale, we still evaluate confidence for reporting
+                # But we don't confirm/reject stale ones automatically unless confidence is extreme
+                # Actually, the task says to check for stale topics and archive if needed. We've done that.
+                # We still need to apply confidence thresholds to non-stale or all pending?
+                # Let's continue to evaluate all pending, including stale, but maybe skip confirm/reject for stale?
+                # The task: "For each pending engram... evaluate... If >0.8 confirm, <0.3 reject, otherwise mark for review"
+                # It doesn't say to skip stale ones. So we'll still apply thresholds.
+                # But we already archived it. We can still confirm/reject it if confidence is extreme.
+                # Let's continue.
+
+            # Apply confidence thresholds
+            if confidence > 0.8:
+                eg.correctness.confirm(by="auto_processor", note=f"Auto-confirmed: confidence {confidence:.2f}")
+                store.save(eg)
+                actions["confirmed"] += 1
+                details.append(f"✅ Confirmed: [{engram_id[:8]}] {content_preview} (conf: {confidence:.2f})")
+            elif confidence < 0.3:
+                eg.correctness.reject(by="auto_processor", note=f"Auto-rejected: confidence {confidence:.2f}")
+                store.save(eg)
+                actions["rejected"] += 1
+                details.append(f"❌ Rejected: [{engram_id[:8]}] {content_preview} (conf: {confidence:.2f})")
+            else:
+                actions["left_for_review"] += 1
+                details.append(f"⏳ Review later: [{engram_id[:8]}] {content_preview} (conf: {confidence:.2f})")
+
+        except Exception as e:
+            actions["errors"] += 1
+            details.append(f"⚠️ Error processing engram: {str(e)}")
+
+    # Generate summary report
+    report_lines = []
+    report_lines.append("=" * 60)
+    report_lines.append("PENDING ENGRAMS PROCESSING REPORT")
+    report_lines.append("=" * 60)
+    report_lines.append(f"Timestamp: {datetime.now(timezone.utc).isoformat()}")
+    report_lines.append(f"Total engrams: {len(all_engrams)}")
+    report_lines.append(f"Pending engrams processed: {len(pending)}")
+    report_lines.append("")
+    report_lines.append("ACTIONS TAKEN:")
+    report_lines.append(f"  ✅ Auto-confirmed (confidence > 0.8): {actions['confirmed']}")
+    report_lines.append(f"  ❌ Auto-rejected (confidence < 0.3): {actions['rejected']}")
+    report_lines.append(f"  ⏳ Left for review (0.3 ≤ confidence ≤ 0.8): {actions['left_for_review']}")
+    report_lines.append(f"  📦 Archived stale topics: {actions['archived_stale']}")
+    report_lines.append(f"  ⚠️ Errors: {actions['errors']}")
+    report_lines.append("")
+    report_lines.append("DETAILS:")
+    report_lines.extend(details)
+    report_lines.append("")
+    report_lines.append("=" * 60)
+
+    report = "\n".join(report_lines)
+
+    # Print to stdout
+    print("\n" + report)
+
+    # Save report to file
+    report_dir = Path(__file__).parent.parent / "reports"
+    report_dir.mkdir(parents=True, exist_ok=True)
+    report_file = report_dir / f"pending_engrams_{datetime.now().strftime('%Y%m%d_%H%M%S')}.txt"
+    report_file.write_text(report, encoding="utf-8")
+    print(f"\n📄 Report saved to: {report_file}")
+
+    store.close()
+    return actions
+
+
+if __name__ == "__main__":
+    result = process_pending_engrams()
+    print("\nProcessing complete.")

scripts/smoke_dashboard.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+HOST="${SECOND_BRAIN_HOST:-127.0.0.1}"
+PORT="${SECOND_BRAIN_PORT:-${PORT:-8501}}"
+BASE_URL="http://${HOST}:${PORT}"
+
+if command -v systemctl >/dev/null 2>&1; then
+  if ! systemctl is-active --quiet openclaw-secondbrain-dashboard.service; then
+    echo "ERROR: openclaw-secondbrain-dashboard.service is not active" >&2
+    exit 1
+  fi
+fi
+
+curl -fsS "${BASE_URL}/healthz" >/dev/null
+
+stats_json="$(curl -fsS "${BASE_URL}/api/stats")"
+python3 -c 'import json,sys; json.load(sys.stdin)' <<<"$stats_json"
+
+echo "OK: dashboard smoke test passed (${BASE_URL})"

src/engram.py

@@ -40,26 +40,60 @@ class ReviewEntry:
 @dataclass
 class Correctness:
     """Verfolgt die Korrektheit eines Engramms über Zeit."""
+    # verdict model (not only binary confirm/reject)
+    # Values:
+    # - unknown
+    # - probable_true / probable_false
+    # - confirmed_true / confirmed_false
+    verdict: str = "unknown"
+    evidence: List[Dict[str, Any]] = field(default_factory=list)
     confirmed: bool = False
     confirmations: int = 0
     rejections: int = 0
     last_reviewed: Optional[str] = None
     review_history: List[ReviewEntry] = field(default_factory=list)
 
+    def is_final(self) -> bool:
+        return self.verdict in ("confirmed_true", "confirmed_false")
+
+    def set_verdict(self, by: str, verdict: str, note: str = "", evidence: Optional[List[Dict[str, Any]]] = None) -> None:
+        verdict = (verdict or "").strip()
+        if verdict not in ("unknown", "probable_true", "probable_false", "confirmed_true", "confirmed_false"):
+            verdict = "unknown"
+        self.verdict = verdict
+        # Keep backward-compatible boolean in sync:
+        # historically, confirmed=True meant "this statement is correct".
+        self.confirmed = verdict == "confirmed_true"
+        self.last_reviewed = _now()
+        if evidence:
+            try:
+                self.evidence.extend([e for e in evidence if isinstance(e, dict)])
+            except Exception:
+                pass
+        self.review_history.append(ReviewEntry(by, "set_verdict", self.last_reviewed, f"{verdict}: {note}".strip()))
+
     def confirm(self, by: str, note: str = "") -> None:
         self.confirmations += 1
-        self.confirmed = True
-        self.last_reviewed = _now()
+        self.set_verdict(by, "confirmed_true", note)
+        # Preserve historic action tag too
         self.review_history.append(ReviewEntry(by, "confirm", self.last_reviewed, note))
 
     def reject(self, by: str, note: str = "") -> None:
         self.rejections += 1
-        self.confirmed = False
-        self.last_reviewed = _now()
+        self.set_verdict(by, "confirmed_false", note)
         self.review_history.append(ReviewEntry(by, "reject", self.last_reviewed, note))
 
     def score(self) -> float:
         """Confidence-Score aus Korrekturhistorie."""
+        # verdict-first scoring (explicit, non-binary)
+        if self.verdict == "confirmed_true":
+            return 1.0
+        if self.verdict == "confirmed_false":
+            return 0.0
+        if self.verdict == "probable_true":
+            return 0.75
+        if self.verdict == "probable_false":
+            return 0.25
         total = self.confirmations + self.rejections
         if total == 0:
             return 0.5  # Unbestimmt
@@ -74,6 +108,8 @@ class Correctness:
             else:
                 review_history.append(entry.to_dict())
         return {
+            "verdict": self.verdict,
+            "evidence": self.evidence,
             "confirmed": self.confirmed,
             "confirmations": self.confirmations,
             "rejections": self.rejections,
@@ -84,11 +120,30 @@ class Correctness:
     @classmethod
     def from_dict(cls, d: dict) -> "Correctness":
         c = cls()
+        verdict = d.get("verdict")
+        if isinstance(verdict, str) and verdict.strip():
+            c.verdict = verdict.strip()
         c.confirmed = d.get("confirmed", False)
         c.confirmations = d.get("confirmations", 0)
         c.rejections = d.get("rejections", 0)
         c.last_reviewed = d.get("last_reviewed")
+        ev = d.get("evidence", [])
+        if isinstance(ev, list):
+            c.evidence = [e for e in ev if isinstance(e, dict)]
         c.review_history = [ReviewEntry.from_dict(r) for r in d.get("review_history", [])]
+        # Backfill verdict if missing/invalid.
+        if c.verdict not in ("unknown", "probable_true", "probable_false", "confirmed_true", "confirmed_false"):
+            if c.confirmed:
+                c.verdict = "confirmed_true"
+            elif c.rejections > 0:
+                c.verdict = "confirmed_false"
+            else:
+                c.verdict = "unknown"
+        # Ensure boolean stays consistent for older mixed data.
+        if c.verdict == "confirmed_true":
+            c.confirmed = True
+        elif c.verdict == "confirmed_false":
+            c.confirmed = False
         return c
 
 

src/graph_view.py

@@ -134,11 +134,19 @@ def _node_size(access_count: int) -> float:
 
 def generate_graph_html(store: EngramStore, output_path: str) -> str:
     """Generiert interaktive HTML-Graph-Visualisierung."""
-    engrams = store.get_all()
+    # store.get_all() defaults to 1000; paginate so the graph can include all nodes.
+    engrams = []
+    offset = 0
+    while True:
+        batch = store.get_all(limit=2000, offset=offset)
+        if not batch:
+            break
+        engrams.extend(batch)
+        offset += len(batch)
 
     nodes = []
     edges = []
-    node_ids = set()
+    node_ids = set(str(e.id) for e in engrams)
 
     for eg in engrams:
         eid = str(eg.id)
@@ -156,22 +164,19 @@ def generate_graph_html(store: EngramStore, output_path: str) -> str:
                 "size": size,
                 "confidence": conf,
                 "confirmed": eg.correctness.confirmed,
+                "verdict": getattr(eg.correctness, "verdict", "unknown"),
                 "source": eg.metadata.get("source", "?"),
                 "tags": tags,
             }
         })
-        node_ids.add(eid)
 
+    # Add edges after all nodes are known (otherwise early nodes miss links).
+    for eg in engrams:
+        eid = str(eg.id)
         for lid in eg.links:
             lid_s = str(lid)
             if lid_s in node_ids:
-                edges.append({
-                    "data": {
-                        "id": f"{eid}_{lid_s}",
-                        "source": eid,
-                        "target": lid_s,
-                    }
-                })
+                edges.append({"data": {"id": f"{eid}_{lid_s}", "source": eid, "target": lid_s}})
 
     elements = {"nodes": nodes, "edges": edges}
     html = _HTML_TEMPLATE.format(elements_json=json.dumps(elements, ensure_ascii=False))

src/openclaw_bridge.py

@@ -12,6 +12,7 @@ import os
 import sys
 import json
 import traceback
+import re
 from pathlib import Path
 from datetime import datetime, timezone
 from typing import Optional, List, Dict, Any
@@ -48,6 +49,45 @@ except Exception:
 # --- Konfiguration ---
 BRAIN_DB = Path(__file__).parent.parent / "data" / "brain.sqlite"
 
+_UUID_RE = re.compile(r"\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b", re.I)
+_SHORT_ID_RE = re.compile(r"\b[0-9a-f]{8}\b", re.I)
+
+
+def _detect_feedback(content: str) -> dict | None:
+    """
+    Heuristik: erkennt kurze Korrektur-/Feedback-Nachrichten in Chats.
+
+    Returns:
+      {"kind":"confirm"|"reject"|"stop", "target": <id or short-id or None>, "raw": <normalized>}
+    """
+    if not isinstance(content, str):
+        return None
+    raw = content.strip()
+    if not raw:
+        return None
+    norm = raw.lower().strip()
+
+    target = None
+    m = _UUID_RE.search(raw)
+    if m:
+        target = m.group(0)
+    else:
+        m2 = _SHORT_ID_RE.search(raw)
+        if m2:
+            target = m2.group(0)
+
+    if norm in {"stop", "stopp", "halt"}:
+        return {"kind": "stop", "target": target, "raw": norm}
+    if norm in {"nein", "no", "falsch", "wrong"}:
+        return {"kind": "reject", "target": target, "raw": norm}
+    if norm in {"ja", "yes", "richtig", "korrekt", "stimmt"}:
+        return {"kind": "confirm", "target": target, "raw": norm}
+    if norm.startswith(("korrigiert", "korrektur", "correction")):
+        if "richtig" in norm or "korrekt" in norm:
+            return {"kind": "confirm", "target": target, "raw": norm}
+        return {"kind": "reject", "target": target, "raw": norm}
+    return None
+
 
 def get_brain() -> EngramStore:
     """Gibt initialisierten Brain-Store."""
@@ -77,14 +117,43 @@ def save_session_learned(
         )
     """
     store = get_brain()
+    tags = tags or []
+
+    fb = _detect_feedback(content) if source == "session" else None
+    fb_target_id: str | None = None
+    if fb:
+        tags = list(dict.fromkeys(tags + ["feedback", fb["kind"]]))
+        confidence = min(confidence, 0.2)
+        grounding = Grounding.ASSUMPTION
+        if session_id:
+            recent = store.get_latest_by_session_id(session_id, limit=10, exclude_tags=["feedback"])
+            if recent:
+                fb_target_id = str(recent[0].id)
+
     eg = Engram.create(
         content=content,
         source=source,
-        tags=tags or [],
+        tags=tags,
         session_id=session_id,
         confidence=confidence,
         grounding=grounding,
     )
+    if fb and fb_target_id:
+        target = store.get(fb_target_id)
+        if target:
+            try:
+                # Link both ways for graphing/traceability
+                eg.links.append(target.id)
+                if eg.id not in target.links:
+                    target.links.append(eg.id)
+
+                # Lock target so auto-review does not keep "re-deciding" after a correction signal.
+                target.metadata["predict_locked"] = True
+                target.metadata["predict_locked_reason"] = f"feedback:{fb['raw']}"
+                target.metadata["predict_locked_at"] = datetime.now(timezone.utc).isoformat()
+                store.save(target)
+            except Exception:
+                pass
     store.save(eg)
     return eg
 

src/store.py

@@ -6,6 +6,7 @@ Keine externen Abhängigkeiten außer sqlite3 (stdlib).
 import json
 import sqlite3
 import os
+import re
 from pathlib import Path
 from typing import List, Optional, Dict, Any
 from uuid import UUID
@@ -135,6 +136,32 @@ class EngramStore:
         ).fetchall()
         return [self._row_to_engram(r) for r in rows]
 
+    def get_latest_by_session_id(
+        self,
+        session_id: str,
+        *,
+        limit: int = 5,
+        exclude_tags: Optional[List[str]] = None,
+    ) -> List[Engram]:
+        """
+        Lädt die neuesten Engramme für eine OpenClaw-Session-ID.
+
+        Hinweis: session_id liegt im `metadata_json`; wir nutzen eine robuste
+        LIKE-Suche, damit auch Legacy-Records gefunden werden.
+        """
+        if not session_id:
+            return []
+        rows = self._conn.execute(
+            "SELECT * FROM engrams WHERE metadata_json LIKE ? ORDER BY created_at DESC LIMIT ?",
+            (f'%"session_id": "{session_id}"%', limit),
+        ).fetchall()
+        engrams = [self._row_to_engram(r) for r in rows]
+        if exclude_tags:
+            ex = set(t for t in exclude_tags if isinstance(t, str))
+            if ex:
+                engrams = [e for e in engrams if not (ex & set(e.metadata.get("tags", []) or []))]
+        return engrams
+
     def delete(self, engram_id: str) -> bool:
         """Löscht ein Engramm und alle Verknüpfungen."""
         rowid = self._conn.execute(
@@ -158,7 +185,12 @@ class EngramStore:
     def search_text(self, query: str, limit: int = 10) -> List[Engram]:
         """Full-Text-Suche über Engramm-Inhalt via SQLite FTS5 (OR-Verknüpfung)."""
         # FTS5-Syntax: Wörter mit OR verbinden für bessere Ergebnisse
-        words = [w.strip() for w in query.replace("'", "''").split() if w.strip()]
+        words = []
+        for word in query.split():
+            # Nur alphanumerische Zeichen als FTS5-Tokens akzeptieren
+            clean_word = re.sub(r'[^a-zA-Z0-9]+', '', word)
+            if clean_word:
+                words.append(clean_word)
         safe_query = " OR ".join(words) if len(words) > 1 else (words[0] if words else "*")
         sql = """
             SELECT e.* FROM engrams e

static/style.css

@@ -27,6 +27,31 @@ body {
     top: 0;
     z-index: 50;
 }
+
+.tabs-bar{
+    display:flex;
+    gap:8px;
+    padding:8px 12px 10px;
+    background:#141419;
+    border-bottom:1px solid #252530;
+    position: sticky;
+    top: 52px;
+    z-index: 45;
+}
+.tabs-bar .tab-btn{
+    flex:1;
+    background:#1e1e28;
+    border:1px solid #2a2a3a;
+    border-radius: 12px;
+    padding:10px 10px;
+    color:#cfd3ff;
+    font-weight:700;
+    font-size:0.82rem;
+}
+.tabs-bar .tab-btn.active{
+    border-color:#6c8af5;
+    box-shadow:0 0 0 1px rgba(108,138,245,0.22) inset;
+}
 .stat {
     text-align: center;
     min-width: 60px;
@@ -54,6 +79,122 @@ body {
     padding: 10px 12px;
     background: #141419;
 }
+
+/* tab buttons styled via .tabs-bar */
+
+/* ─── Panels (Graph/Status) ──────────────────────────────────────────────── */
+.panel {
+    margin: 8px 12px;
+    background: #1a1a24;
+    border: 1px solid #252533;
+    border-radius: 14px;
+    padding: 10px 12px;
+}
+.panel-title {
+    font-size: 0.8rem;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+    color: #9aa3d9;
+    margin-bottom: 8px;
+}
+.kv-row {
+    display: flex;
+    gap: 10px;
+    padding: 6px 0;
+    border-bottom: 1px solid #20202a;
+}
+.kv-row:last-child { border-bottom: none; }
+.kv-key {
+    width: 110px;
+    color: #888899;
+    font-size: 0.78rem;
+}
+.kv-val {
+    flex: 1;
+    color: #e8e8ee;
+    font-size: 0.85rem;
+    word-break: break-word;
+}
+.pill {
+    display: inline-block;
+    margin: 2px 4px 2px 0;
+    padding: 2px 8px;
+    border-radius: 999px;
+    background: #2a2a3a;
+    color: #8a9aff;
+    font-size: 0.72rem;
+}
+
+.verdict-pill{
+    display:inline-block;
+    margin: 2px 6px 2px 0;
+    padding: 2px 8px;
+    border-radius: 999px;
+    font-size: 0.72rem;
+    font-weight: 800;
+    letter-spacing: 0.4px;
+    border: 1px solid #2a2a3a;
+    background: #1e1e28;
+    color: #cfd3ff;
+}
+.verdict-pill.v-true{ border-color:#2f6b3f; color:#aaf0b6; }
+.verdict-pill.v-false{ border-color:#7a2c2c; color:#ffb3b3; }
+.verdict-pill.v-prob-true{ border-color:#6c8af5; color:#cfd9ff; }
+.verdict-pill.v-prob-false{ border-color:#b08a2a; color:#ffe2a3; }
+.verdict-pill.v-unknown{ border-color:#3a3a55; color:#b9b9c9; }
+.muted {
+    color: #888899;
+    font-size: 0.8rem;
+    margin-top: 6px;
+}
+.small { font-size: 0.75rem; }
+
+/* Graph canvas */
+#graphCanvas{
+    display:block;
+    margin: 8px auto 0;
+    background:#12121a;
+    border:1px solid #252533;
+    border-radius: 14px;
+    touch-action: none;
+}
+
+.graph-controls{
+    display:flex;
+    gap:8px;
+    padding: 10px 12px 0;
+    align-items:center;
+    flex-wrap: wrap;
+}
+.graph-controls .btn{
+    background:#1e1e28;
+    border:1px solid #2a2a3a;
+    border-radius: 10px;
+    padding: 8px 10px;
+    color:#cfd3ff;
+    font-weight:700;
+    font-size:0.82rem;
+}
+.graph-controls .btn.primary{
+    border-color:#6c8af5;
+    box-shadow:0 0 0 1px rgba(108,138,245,0.18) inset;
+}
+.graph-legend{
+    margin: 8px 12px 0;
+    padding: 10px 12px;
+    background:#1a1a24;
+    border:1px solid #252533;
+    border-radius: 14px;
+    color:#b9b9c9;
+    font-size:0.8rem;
+    line-height:1.4;
+}
+.legend-row{ display:flex; align-items:center; gap:8px; margin-top:6px; }
+.legend-dot{ width:10px; height:10px; border-radius:50%; display:inline-block; }
+.legend-dot.engram{ background:#6c8af5; }
+.legend-dot.tag{ background:#8a9aff; }
+.legend-dot.match{ background:#f7d154; }
+.graph-hint{ padding: 4px 12px 10px; }
 #searchInput {
     flex: 1;
     background: #1e1e28;

systemd/openclaw-memory-archive.service

@@ -0,0 +1,9 @@
+[Unit]
+Description=OpenClaw archive memory/*.md older than 7 days
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py archive_memory_md'
+

systemd/openclaw-memory-archive.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw archive memory/*.md daily
+
+[Timer]
+OnCalendar=*-*-* 03:30:00
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-backup.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain backup_secondbrain
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py backup_secondbrain'

systemd/openclaw-secondbrain-backup.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain backup_secondbrain (daily 02:00)
+
+[Timer]
+OnCalendar=*-*-* 02:00:00
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-dashboard.service

@@ -0,0 +1,15 @@
+[Unit]
+Description=OpenClaw Second-Brain Dashboard (FastAPI)
+After=network.target
+
+[Service]
+Type=simple
+WorkingDirectory=/root/.openclaw/workspace/second-brain
+Environment=SECOND_BRAIN_WORKSPACE=/root/.openclaw/workspace/second-brain
+Environment=SECOND_BRAIN_PORT=8501
+ExecStart=/root/.openclaw/workspace/second-brain/.venv/bin/uvicorn fastapi_app:app --host 0.0.0.0 --port 8501
+Restart=always
+RestartSec=2
+
+[Install]
+WantedBy=multi-user.target

systemd/openclaw-secondbrain-export-obsidian.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain export_obsidian
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py export_obsidian'

systemd/openclaw-secondbrain-export-obsidian.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain export_obsidian (hourly)
+
+[Timer]
+OnCalendar=hourly
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-heartbeat.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain heartbeat_secondbrain
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py heartbeat_secondbrain'

systemd/openclaw-secondbrain-heartbeat.timer

@@ -0,0 +1,11 @@
+[Unit]
+Description=OpenClaw Second-Brain heartbeat_secondbrain (every 6h)
+
+[Timer]
+OnCalendar=*-*-* 00,06,12,18:00:00
+RandomizedDelaySec=300
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-index-vectors.service

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain index_vectors
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+Environment=HF_HOME=/root/.openclaw/workspace/second-brain/data/hf_cache
+Environment=SENTENCE_TRANSFORMERS_HOME=/root/.openclaw/workspace/second-brain/data/st_cache
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py index_vectors'

systemd/openclaw-secondbrain-index-vectors.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain index_vectors (every 30 min)
+
+[Timer]
+OnCalendar=*:0/30
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-ingest-memory.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest_memory
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py ingest_memory'

systemd/openclaw-secondbrain-ingest-memory.timer

@@ -0,0 +1,9 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest_memory (every 5 min)
+
+[Timer]
+OnCalendar=*:0/5
+Persistent=true
+
+[Install]
+WantedBy=timers.target

systemd/openclaw-secondbrain-ingest-obsidian.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest_obsidian
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py ingest_obsidian'

systemd/openclaw-secondbrain-ingest-obsidian.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest_obsidian (every 15 min)
+
+[Timer]
+OnCalendar=*:0/15
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-ingest-transcript-to-db.service

@@ -0,0 +1,9 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest transcript -> DB
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py ingest_transcript_to_db'
+

systemd/openclaw-secondbrain-ingest-transcript-to-db.timer

@@ -0,0 +1,12 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest transcript -> DB (every 5 min)
+
+[Timer]
+OnBootSec=90s
+OnUnitActiveSec=300s
+Unit=openclaw-secondbrain-ingest-transcript-to-db.service
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-ingest-transcript-to-memory.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest transcript -> memory/*.md
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/second-brain/cron_tasks/ingest_transcript_to_memory.py'

systemd/openclaw-secondbrain-ingest-transcript-to-memory.timer

@@ -0,0 +1,11 @@
+[Unit]
+Description=OpenClaw Second-Brain ingest transcript -> memory (every 1 min)
+
+[Timer]
+OnBootSec=30s
+OnUnitActiveSec=60s
+Unit=openclaw-secondbrain-ingest-transcript-to-memory.service
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-notify@.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain failure notify (%i)
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc '/root/.openclaw/workspace/notify-telegram.sh "❌ Second-Brain job failed: %i. Check: journalctl -u %i -n 50 --no-pager"'
+

systemd/openclaw-secondbrain-proactive-search.service

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain proactive_search_wrapper
+Wants=network-online.target
+After=network-online.target
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py proactive_search_wrapper'

systemd/openclaw-secondbrain-proactive-search.timer

@@ -0,0 +1,11 @@
+[Unit]
+Description=OpenClaw Second-Brain proactive_search_wrapper (every 4h)
+
+[Timer]
+OnCalendar=*-*-* 01,05,09,13,17,21:10:00
+RandomizedDelaySec=600
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+

systemd/openclaw-secondbrain-review.service

@@ -0,0 +1,8 @@
+[Unit]
+Description=OpenClaw Second-Brain review_brain
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py review_brain'

systemd/openclaw-secondbrain-review.timer

@@ -0,0 +1,9 @@
+[Unit]
+Description=OpenClaw Second-Brain review_brain (every 5 min)
+
+[Timer]
+OnCalendar=*:0/5
+Persistent=true
+
+[Install]
+WantedBy=timers.target

systemd/openclaw-secondbrain-task@.service

@@ -0,0 +1,13 @@
+[Unit]
+Description=OpenClaw Second-Brain task (%i)
+Wants=network-online.target
+After=network-online.target
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py %i
+Nice=10
+IOSchedulingClass=best-effort
+IOSchedulingPriority=6
+

systemd/openclaw-secondbrain-verify-pending.service

@@ -0,0 +1,9 @@
+[Unit]
+Description=OpenClaw Second-Brain verify_pending_external
+OnFailure=openclaw-secondbrain-notify@%n.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/root/.openclaw/workspace
+ExecStart=/bin/bash -lc 'flock -n /tmp/%n.lock /usr/bin/python3 /root/.openclaw/workspace/openclaw_cron_wrapper.py verify_pending_external'
+

systemd/openclaw-secondbrain-verify-pending.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=OpenClaw Second-Brain periodic verify_pending_external
+
+[Timer]
+OnBootSec=2min
+OnUnitActiveSec=5min
+Persistent=true
+
+[Install]
+WantedBy=timers.target

templates/dashboard.html

@@ -16,6 +16,12 @@
             <div class="stat"><span class="stat-num" id="statErrors">-</span><span class="stat-label">Err</span></div>
         </header>
 
+        <div class="tabs-bar">
+            <button class="tab-btn active" id="tabCards" onclick="setView('cards')">Cards</button>
+            <button class="tab-btn" id="tabGraph" onclick="setView('graph')">Graph</button>
+            <button class="tab-btn" id="tabStatus" onclick="setView('status')">Status</button>
+        </div>
+
         <!-- Search -->
         <div class="search-box">
             <input type="text" id="searchInput" placeholder="🔍 Suche..." />
@@ -23,6 +29,7 @@
                 <option value="all">Alle</option>
                 <option value="pending">Pending</option>
                 <option value="confirmed">Confirmed</option>
+                <option value="rejected">Rejected</option>
                 <option value="errors">Errors</option>
             </select>
         </div>
@@ -37,6 +44,38 @@
         <!-- Cards -->
         <div class="cards" id="cards"></div>
 
+        <!-- Graph -->
+	        <div class="graph" id="graph" style="display:none;">
+	            <div class="graph-controls">
+	                <button class="btn primary" id="btnGraphPhysics" onclick="toggleGraphPhysics()">Physics: off</button>
+	                <button class="btn" onclick="resetGraphView()">Reset view</button>
+	                <button class="btn" onclick="fitGraphView()">Fit</button>
+	                <label class="muted small" style="display:flex;align-items:center;gap:8px">
+	                    <span>Physics</span>
+	                    <input id="physicsStrength" type="range" min="0" max="100" value="60" oninput="setPhysicsStrength(this.value)" style="width:140px">
+	                    <span id="physicsStrengthVal">60</span>
+	                </label>
+	                <select class="btn" id="graphLimit" onchange="reloadGraph()" title="Wie viele Knoten laden? 0=all">
+	                    <option value="0">Nodes: all</option>
+	                    <option value="200">Nodes: 200</option>
+	                    <option value="1000">Nodes: 1000</option>
+	                    <option value="5000">Nodes: 5000</option>
+	                </select>
+                <button class="btn" onclick="reloadGraph()">Reload</button>
+            </div>
+            <canvas id="graphCanvas" width="440" height="520"></canvas>
+            <div class="graph-hint muted small" id="graphHint">Lade Graph…</div>
+            <div class="graph-legend">
+                <div><strong>Graph</strong>: Zoom per Pinch (2 Finger), Pan per Drag (1 Finger). Tap auf Engram öffnet Details, Tap auf Tag setzt Suche.</div>
+                <div class="legend-row"><span class="legend-dot engram"></span> Engram</div>
+                <div class="legend-row"><span class="legend-dot tag"></span> Tag</div>
+                <div class="legend-row"><span class="legend-dot match"></span> Match (Suche)</div>
+            </div>
+        </div>
+
+        <!-- Status -->
+        <div class="status" id="status" style="display:none;"></div>
+
         <!-- Pagination -->
         <div class="pagination" id="pagination">
             <button id="btnPrev" onclick="prevPage()">◀</button>
@@ -67,12 +106,25 @@ let state = {
     filter: 'all',
     search: '',
     autoRefresh: true,
+    view: 'cards',
+    lastEvent: null,
 };
 
 // ─── Fetch ──────────────────────────────────────────────────────────────────
 async function api(path, opts = {}) {
     const r = await fetch(path, opts);
-    if (!r.ok) throw new Error((await r.json()).error || r.statusText);
+    if (!r.ok) {
+        let msg = r.statusText;
+        try {
+            const j = await r.json();
+            msg = j.error || j.detail || msg;
+            if (Array.isArray(msg)) msg = JSON.stringify(msg);
+            if (typeof msg !== 'string') msg = JSON.stringify(msg);
+        } catch (e) {
+            // ignore JSON parse errors
+        }
+        throw new Error(msg);
+    }
     return r.json();
 }
 
@@ -84,11 +136,36 @@ async function loadStats() {
     document.getElementById('statErrors').textContent = s.errors;
 }
 
+function updateStatsFromEvent(ev) {
+    if (!ev || !ev.stats) return;
+    const s = ev.stats;
+    document.getElementById('statTotal').textContent = s.total;
+    document.getElementById('statConfirmed').textContent = s.confirmed;
+    document.getElementById('statPending').textContent = s.pending;
+    document.getElementById('statErrors').textContent = s.errors;
+}
+
+function setView(view) {
+    state.view = view;
+    document.getElementById('tabCards').classList.toggle('active', view === 'cards');
+    document.getElementById('tabGraph').classList.toggle('active', view === 'graph');
+    document.getElementById('tabStatus').classList.toggle('active', view === 'status');
+
+    document.getElementById('cards').style.display = view === 'cards' ? '' : 'none';
+    document.getElementById('pagination').style.display = view === 'cards' ? '' : 'none';
+    document.getElementById('graph').style.display = view === 'graph' ? '' : 'none';
+    document.getElementById('status').style.display = view === 'status' ? '' : 'none';
+
+    if (view === 'graph') loadGraph();
+    if (view === 'status') loadStatus();
+}
+
 async function loadCards() {
     let url = `/api/engrams?limit=${state.limit}&offset=${state.offset}`;
     if (state.search) url += `&search=${encodeURIComponent(state.search)}`;
     if (state.filter === 'confirmed') url += '&confirmed=1';
     if (state.filter === 'pending') url += '&confirmed=0';
+    if (state.filter === 'rejected') url += '&verdict=confirmed_false';
     if (state.filter === 'errors') url += '&tag=error';
 
     const data = await api(url);
@@ -99,12 +176,788 @@ async function loadCards() {
     document.getElementById('btnNext').disabled = data.items.length < state.limit;
 }
 
+async function loadStatus() {
+    const reqs = await Promise.allSettled([
+        api('/api/config'),
+        api('/api/db_info'),
+        api('/api/jobs'),
+        api('/api/insights?limit=8'),
+        api('/api/storage_stats'),
+        api('/api/pending?limit=20&offset=0'),
+    ]);
+    const pick = (i, fallback) => (reqs[i].status === 'fulfilled' ? reqs[i].value : fallback);
+    const err = (i) => (reqs[i].status === 'rejected' ? (reqs[i].reason && reqs[i].reason.message ? reqs[i].reason.message : String(reqs[i].reason)) : null);
+
+    const cfg = pick(0, { workspace: '-', db_path: '-' });
+    const db = pick(1, { db_path: '-', mtime: null });
+    const jobs = pick(2, { items: [], error: err(2) });
+    const ins = pick(3, { pending: '-', top_tags: [], top_hosts: [], error: err(3) });
+    const stor = pick(4, { sql: { total_engrams: '-', confirmed: '-', pending: '-', by_source: {} }, vector: { chroma_size_bytes: 0, embedding_cache_files: 0 }, obsidian: { configured: false } });
+    const pend = pick(5, { items: [], error: err(5) });
+
+    const el = document.getElementById('status');
+    const jobsHtml = (jobs.items || []).map(j => `
+        <div class="kv-row">
+            <div class="kv-key">${j.unit}</div>
+            <div class="kv-val">${j.error ? ('ERR: ' + j.error) : (j.active + '/' + j.sub)}</div>
+        </div>
+    `).join('');
+
+    const topTags = (ins.top_tags || []).map(t => `<span class="pill">${t.key} (${t.count})</span>`).join(' ');
+    const topHosts = (ins.top_hosts || []).map(t => `<span class="pill">${t.key} (${t.count})</span>`).join(' ');
+    const bySource = Object.entries((stor.sql && stor.sql.by_source) ? stor.sql.by_source : {})
+        .slice(0, 8)
+        .map(([k,v]) => `<span class="pill">${k}: ${v}</span>`)
+        .join(' ');
+
+    const pendItems = (pend.items || []);
+    const pendHtml = pendItems.map(p => `
+        <div class="kv-row" onclick="showDetail('${p.id}')">
+            <div class="kv-key">${(p.source||'').slice(0,12)}</div>
+            <div class="kv-val">
+                <span class="pill">${p.id.substring(0,8)}</span>
+                ${escapeHtml((p.content||'').substring(0,120))}${(p.content||'').length>120?'…':''}
+                <div class="actions" style="margin-top:6px" onclick="event.stopPropagation()">
+                    <button class="btn-ok" onclick="confirm('${p.id}', event)">✅</button>
+                    <button class="btn-no" onclick="reject('${p.id}', event)">❌</button>
+                </div>
+            </div>
+        </div>
+    `).join('');
+
+    el.innerHTML = `
+        <div class="panel">
+            <div class="panel-title">Config</div>
+            <div class="kv-row"><div class="kv-key">workspace</div><div class="kv-val">${cfg.workspace}</div></div>
+            <div class="kv-row"><div class="kv-key">db</div><div class="kv-val">${db.db_path}</div></div>
+            <div class="kv-row"><div class="kv-key">db mtime</div><div class="kv-val">${new Date(db.mtime).toLocaleString()}</div></div>
+        </div>
+        <div class="panel">
+            <div class="panel-title">Storage</div>
+            <div class="kv-row"><div class="kv-key">SQL</div><div class="kv-val">${stor.sql.total_engrams} engrams (ok ${stor.sql.confirmed}, pending ${stor.sql.pending})</div></div>
+            <div class="kv-row"><div class="kv-key">Vector</div><div class="kv-val">chroma ${(stor.vector.chroma_size_bytes/1024/1024).toFixed(1)} MB, cache ${stor.vector.embedding_cache_files} files</div></div>
+            <div class="kv-row"><div class="kv-key">Obsidian</div><div class="kv-val">${stor.obsidian.configured ? 'configured' : 'not configured'}</div></div>
+            <div class="kv-row"><div class="kv-key">By source</div><div class="kv-val">${bySource || '-'}</div></div>
+        </div>
+        <div class="panel">
+            <div class="panel-title">Jobs</div>
+            ${jobs.error ? `<div class="muted">Fehler: ${escapeHtml(jobs.error)}</div>` : (jobsHtml || '<div class="muted">Keine Daten</div>')}
+        </div>
+        <div class="panel">
+            <div class="panel-title">Insights</div>
+            ${ins.error ? `<div class="muted">Fehler: ${escapeHtml(ins.error)}</div>` : ''}
+            <div class="kv-row"><div class="kv-key">pending</div><div class="kv-val">${ins.pending}</div></div>
+            <div class="kv-row"><div class="kv-key">top tags</div><div class="kv-val">${topTags || '-'}</div></div>
+            <div class="kv-row"><div class="kv-key">top hosts</div><div class="kv-val">${topHosts || '-'}</div></div>
+        </div>
+        <div class="panel">
+            <div class="panel-title">Pending Queue (latest)</div>
+            ${pend.error ? `<div class="muted">Fehler: ${escapeHtml(pend.error)}</div>` : (pendHtml || '<div class="muted">Keine Pendings</div>')}
+        </div>
+    `;
+}
+
+async function loadGraph() {
+    const sel = document.getElementById('graphLimit');
+    const fromSel = sel ? parseInt(sel.value || '0', 10) : NaN;
+    const fromStore = parseInt(localStorage.getItem('graphLimit') || '0', 10);
+    const q = (!Number.isNaN(fromSel)) ? fromSel : (Number.isNaN(fromStore) ? 0 : fromStore);
+    if (sel) sel.value = String(q);
+    if (sel) localStorage.setItem('graphLimit', String(q));
+    const hint = document.getElementById('graphHint');
+    if (hint) hint.textContent = 'Lade Graph…';
+    try {
+        const g = await api(`/api/graph?limit_nodes=${q}`);
+        renderGraph(g.nodes || [], g.edges || []);
+    } catch (e) {
+        if (hint) hint.textContent = `Graph-Fehler: ${e && e.message ? e.message : String(e)}`;
+        const canvas = _graphCanvas();
+        const ctx = _graphCtx();
+        if (canvas && ctx) ctx.clearRect(0,0,canvas.width,canvas.height);
+    }
+}
+
+function reloadGraph() { loadGraph(); }
+
+// ─── Graph Renderer (Canvas) ────────────────────────────────────────────────
+let graphState = {
+    nodes: [],
+    edges: [],
+    sim: [],
+    links: [],
+    nodeById: new Map(),
+    simById: new Map(),
+    degree: new Map(),
+    physicsOn: false,
+    draggingId: null,
+    selectedId: null,
+    panning: false,
+    lastX: 0,
+    lastY: 0,
+    panX: 0,
+    panY: 0,
+    zoom: 1,
+    raf: null,
+    search: '',
+    pointers: new Map(),
+    pinchStartDist: null,
+    pinchStartZoom: null,
+    pinchStartPan: null,
+    down: null, // {pointerId, cx, cy, t}
+    physicsStrength: parseInt(localStorage.getItem('physicsStrength') || '60', 10),
+};
+
+function _graphCanvas() { return document.getElementById('graphCanvas'); }
+function _graphCtx() { return _graphCanvas().getContext('2d'); }
+
+function _graphNodeRadius(n) {
+    const d = graphState.degree.get(n.id) || 0;
+    const base = n.kind === 'tag' ? 4 : (n.kind === 'host' ? 5 : 7);
+    const w = (n.weight || 0);
+    const bonus = Math.min(6, Math.sqrt(Math.max(0, w)) * 0.8);
+    return Math.max(3, Math.min(18, base + Math.sqrt(d) + bonus));
+}
+
+function _graphNodeFill(n) {
+    const d = graphState.degree.get(n.id) || 0;
+    const t = Math.max(0, Math.min(0.55, d / 18)); // higher degree -> brighter
+    const mix = (rgb) => rgb.map(c => Math.round(c + (255 - c) * t));
+
+    if (n.kind === 'tag') {
+        const [r,g,b] = mix([167, 139, 250]);
+        return `rgb(${r},${g},${b})`;
+    }
+    if (n.kind === 'host') {
+        const [r,g,b] = mix([245, 158, 11]);
+        return `rgb(${r},${g},${b})`;
+    }
+
+    const verdict = (n.verdict || '').toString();
+    let base = [96, 165, 250]; // pending/unknown = blue
+    if (verdict === 'confirmed_true') base = [34, 197, 94]; // green
+    else if (verdict === 'confirmed_false') base = [239, 68, 68]; // red
+    else if (verdict === 'probable_true') base = [52, 211, 153]; // teal
+    else if (verdict === 'probable_false') base = [251, 191, 36]; // amber
+
+    // Recency brightening (brand new pops)
+    const now = Date.now();
+    const created = Date.parse(n.created || '') || 0;
+    const ageMin = created ? (now - created) / 60000 : 999999;
+    const rec = Math.max(0, Math.min(0.35, (30 - ageMin) / 30 * 0.35));
+    const bump = (c) => Math.round(c + (255 - c) * rec);
+    const [r,g,b] = mix(base).map(bump);
+    return `rgb(${r},${g},${b})`;
+}
+
+function _graphMatches(n, term) {
+    const t = (term || '').trim().toLowerCase();
+    if (!t) return false;
+    const id = (n.id || '').toLowerCase();
+    const label = (n.label || '').toLowerCase();
+    return id.includes(t) || label.includes(t);
+}
+
+function _graphWorldFromScreen(cx, cy) {
+    return {
+        x: (cx - graphState.panX) / graphState.zoom,
+        y: (cy - graphState.panY) / graphState.zoom,
+    };
+}
+
+function _graphHitTest(wx, wy) {
+    for (let i = graphState.sim.length - 1; i >= 0; i--) {
+        const n = graphState.sim[i];
+        const r = _graphNodeRadius(n) + 2;
+        const dx = wx - n.x;
+        const dy = wy - n.y;
+        if ((dx*dx + dy*dy) <= r*r) return n;
+    }
+    return null;
+}
+
+function _graphInitInteractions() {
+    const canvas = _graphCanvas();
+    if (canvas._graphBound) return;
+    canvas._graphBound = true;
+
+    const toCanvasXY = (ev) => {
+        const rect = canvas.getBoundingClientRect();
+        return { cx: ev.clientX - rect.left, cy: ev.clientY - rect.top };
+    };
+
+    const pinchMidpoint = () => {
+        const pts = Array.from(graphState.pointers.values());
+        if (pts.length < 2) return null;
+        return { cx: (pts[0].cx + pts[1].cx) / 2, cy: (pts[0].cy + pts[1].cy) / 2 };
+    };
+
+    const pinchDistance = () => {
+        const pts = Array.from(graphState.pointers.values());
+        if (pts.length < 2) return null;
+        const dx = pts[0].cx - pts[1].cx;
+        const dy = pts[0].cy - pts[1].cy;
+        return Math.sqrt(dx*dx + dy*dy);
+    };
+
+    canvas.addEventListener('pointerdown', (ev) => {
+        canvas.setPointerCapture(ev.pointerId);
+        const pos = toCanvasXY(ev);
+        graphState.pointers.set(ev.pointerId, pos);
+        graphState.down = { pointerId: ev.pointerId, cx: pos.cx, cy: pos.cy, t: Date.now() };
+
+        // When 2 pointers: start pinch
+        if (graphState.pointers.size === 2) {
+            graphState.panning = false;
+            graphState.draggingId = null;
+            graphState.down = null;
+            graphState.pinchStartDist = pinchDistance();
+            graphState.pinchStartZoom = graphState.zoom;
+            graphState.pinchStartPan = { panX: graphState.panX, panY: graphState.panY };
+            _graphDraw();
+            return;
+        }
+
+        // Single pointer: pan or drag node
+        const { cx, cy } = pos;
+        const w = _graphWorldFromScreen(cx, cy);
+        const hit = _graphHitTest(w.x, w.y);
+        graphState.lastX = cx; graphState.lastY = cy;
+        if (hit) {
+            graphState.draggingId = hit.id;
+            graphState.selectedId = hit.id;
+        } else {
+            graphState.panning = true;
+            graphState.selectedId = null;
+        }
+        _graphDraw();
+    });
+
+    canvas.addEventListener('pointermove', (ev) => {
+        if (!graphState.pointers.has(ev.pointerId)) return;
+        const pos = toCanvasXY(ev);
+        graphState.pointers.set(ev.pointerId, pos);
+
+        // Pinch zoom (2 pointers)
+        if (graphState.pointers.size >= 2 && graphState.pinchStartDist) {
+            const dist = pinchDistance();
+            const mid = pinchMidpoint();
+            if (!dist || !mid) return;
+
+            const zoomOld = graphState.zoom;
+            const wx = (mid.cx - graphState.panX) / zoomOld;
+            const wy = (mid.cy - graphState.panY) / zoomOld;
+
+            const factor = dist / graphState.pinchStartDist;
+            const next = Math.max(0.35, Math.min(3.0, graphState.pinchStartZoom * factor));
+            graphState.zoom = next;
+            graphState.panX = mid.cx - wx * graphState.zoom;
+            graphState.panY = mid.cy - wy * graphState.zoom;
+            _graphDraw();
+            return;
+        }
+
+        // Single pointer
+        const { cx, cy } = pos;
+        const dx = cx - graphState.lastX;
+        const dy = cy - graphState.lastY;
+        graphState.lastX = cx; graphState.lastY = cy;
+
+        if (graphState.draggingId) {
+            const w = _graphWorldFromScreen(cx, cy);
+            const n = graphState.simById.get(graphState.draggingId);
+            if (n) { n.x = w.x; n.y = w.y; n.vx = 0; n.vy = 0; }
+            _graphDraw();
+            return;
+        }
+        if (graphState.panning) {
+            graphState.panX += dx;
+            graphState.panY += dy;
+            _graphDraw();
+        }
+    });
+
+    const endPointer = (ev) => {
+        if (graphState.pointers.has(ev.pointerId)) graphState.pointers.delete(ev.pointerId);
+        try { canvas.releasePointerCapture(ev.pointerId); } catch {}
+
+        if (graphState.pointers.size < 2) {
+            graphState.pinchStartDist = null;
+            graphState.pinchStartZoom = null;
+            graphState.pinchStartPan = null;
+        }
+
+        const pos = toCanvasXY(ev);
+        const down = graphState.down && graphState.down.pointerId === ev.pointerId ? graphState.down : null;
+        graphState.down = null;
+
+        const moved = down ? Math.hypot(pos.cx - down.cx, pos.cy - down.cy) : 999;
+        const isTap = !!down && moved <= 8 && (Date.now() - down.t) <= 500;
+
+        const w = _graphWorldFromScreen(pos.cx, pos.cy);
+        const hit = _graphHitTest(w.x, w.y);
+
+        graphState.draggingId = null;
+        graphState.panning = false;
+
+        if (isTap && hit) {
+            graphState.selectedId = hit.id;
+            _graphDraw();
+            if (hit.kind === 'engram') {
+                showDetail(hit.id);
+            } else if (hit.kind === 'tag') {
+                const label = (hit.label || '').replace(/^tag:/, '');
+                const inp = document.getElementById('searchInput');
+                inp.value = label;
+                inp.dispatchEvent(new Event('input'));
+            }
+            return;
+        }
+        _graphDraw();
+    };
+    canvas.addEventListener('pointerup', endPointer);
+    canvas.addEventListener('pointercancel', endPointer);
+}
+
+function renderGraph(nodes, edges) {
+    const canvas = _graphCanvas();
+    const hint = document.getElementById('graphHint');
+    const ctx = _graphCtx();
+
+    // sync physics slider
+    const slider = document.getElementById('physicsStrength');
+    const sliderVal = document.getElementById('physicsStrengthVal');
+    const s = Math.max(0, Math.min(100, parseInt(graphState.physicsStrength || 60, 10)));
+    graphState.physicsStrength = s;
+    if (slider) slider.value = String(s);
+    if (sliderVal) sliderVal.textContent = String(s);
+
+    const w = canvas.parentElement.clientWidth - 24;
+    canvas.width = Math.max(320, w);
+    canvas.height = Math.max(520, Math.min(900, (window.innerHeight || 900) - 260));
+
+    graphState.nodes = nodes || [];
+    graphState.edges = edges || [];
+    graphState.nodeById = new Map(graphState.nodes.map(n => [n.id, n]));
+    graphState.degree = new Map();
+    for (const e of graphState.edges) {
+        graphState.degree.set(e.from, (graphState.degree.get(e.from) || 0) + 1);
+        graphState.degree.set(e.to, (graphState.degree.get(e.to) || 0) + 1);
+    }
+
+    if (!graphState.nodes.length) {
+        hint.textContent = 'Graph: keine Nodes.';
+        ctx.clearRect(0,0,canvas.width,canvas.height);
+        return;
+    }
+
+    hint.textContent = `nodes=${graphState.nodes.length} edges=${graphState.edges.length}`;
+
+    // Build adjacency for deterministic layout (O(n+e))
+    const adj = new Map();
+    const addAdj = (a, b) => {
+        if (!adj.has(a)) adj.set(a, []);
+        adj.get(a).push(b);
+    };
+    for (const e of graphState.edges) {
+        addAdj(e.from, e.to);
+        addAdj(e.to, e.from);
+    }
+
+    const degree = graphState.degree;
+    const visited = new Set();
+    const comps = [];
+    for (const n of graphState.nodes) {
+        if (visited.has(n.id)) continue;
+        const q = [n.id];
+        visited.add(n.id);
+        const comp = [];
+        while (q.length) {
+            const cur = q.pop();
+            comp.push(cur);
+            const neigh = adj.get(cur) || [];
+            for (const nb of neigh) {
+                if (!visited.has(nb)) {
+                    visited.add(nb);
+                    q.push(nb);
+                }
+            }
+        }
+        comps.push(comp);
+    }
+    comps.sort((a, b) => b.length - a.length);
+
+    // component centers in a loose grid/spiral
+    const centers = [];
+    const gap = 240;
+    const cols = Math.max(1, Math.floor(canvas.width / gap));
+    for (let i = 0; i < comps.length; i++) {
+        const cx = (i % cols) * gap - ((cols - 1) * gap) / 2;
+        const cy = Math.floor(i / cols) * gap - (Math.floor((comps.length - 1) / cols) * gap) / 2;
+        centers.push({ cx, cy });
+    }
+
+    const pos = new Map();
+    const R = 46;
+    for (let ci = 0; ci < comps.length; ci++) {
+        const comp = comps[ci];
+        const center = centers[ci] || { cx: 0, cy: 0 };
+
+        // hub: highest degree engram preferred
+        let hub = comp[0];
+        let hubScore = -1;
+        for (const id of comp) {
+            const n = graphState.nodeById.get(id) || {};
+            const score = (degree.get(id) || 0) + (n.kind === 'engram' ? 3 : 0);
+            if (score > hubScore) { hubScore = score; hub = id; }
+        }
+
+        // BFS layers from hub
+        const layer = new Map();
+        layer.set(hub, 0);
+        const qq = [hub];
+        while (qq.length) {
+            const cur = qq.shift();
+            const l = layer.get(cur) || 0;
+            const neigh = adj.get(cur) || [];
+            for (const nb of neigh) {
+                if (!layer.has(nb)) {
+                    layer.set(nb, l + 1);
+                    qq.push(nb);
+                }
+            }
+        }
+
+        const buckets = new Map();
+        for (const id of comp) {
+            const l = layer.get(id);
+            const ll = (typeof l === 'number') ? l : 99;
+            if (!buckets.has(ll)) buckets.set(ll, []);
+            buckets.get(ll).push(id);
+        }
+
+        // Place hubs first, then rings
+        pos.set(hub, { x: center.cx, y: center.cy });
+        const maxLayer = Math.max(...Array.from(buckets.keys()));
+        for (let l = 1; l <= maxLayer; l++) {
+            const ids = buckets.get(l) || [];
+            if (!ids.length) continue;
+            // Stable order: engrams first, then by degree desc
+            ids.sort((a, b) => {
+                const na = graphState.nodeById.get(a) || {};
+                const nb = graphState.nodeById.get(b) || {};
+                const ka = na.kind === 'engram' ? 0 : (na.kind === 'tag' ? 1 : 2);
+                const kb = nb.kind === 'engram' ? 0 : (nb.kind === 'tag' ? 1 : 2);
+                if (ka !== kb) return ka - kb;
+                return (degree.get(b) || 0) - (degree.get(a) || 0);
+            });
+            const rad = l * R;
+            for (let i = 0; i < ids.length; i++) {
+                const ang = (i / ids.length) * Math.PI * 2;
+                pos.set(ids[i], {
+                    x: center.cx + Math.cos(ang) * rad,
+                    y: center.cy + Math.sin(ang) * rad,
+                });
+            }
+        }
+    }
+
+    graphState.sim = graphState.nodes.map(n => {
+        const p = pos.get(n.id) || { x: (Math.random() - 0.5) * canvas.width, y: (Math.random() - 0.5) * canvas.height };
+        return {
+            id: n.id,
+            kind: n.kind,
+            label: n.label || n.id,
+            weight: n.weight,
+            verdict: n.verdict,
+            confidence: n.confidence,
+            created: n.created,
+            modified: n.modified,
+            last_accessed: n.last_accessed,
+            predict_locked: n.predict_locked,
+            x: p.x,
+            y: p.y,
+            vx: 0, vy: 0,
+        };
+    });
+    graphState.simById = new Map(graphState.sim.map(n => [n.id, n]));
+    graphState.links = graphState.edges
+        .map(e => ({a: graphState.simById.get(e.from), b: graphState.simById.get(e.to), kind: e.kind, weight: e.weight || 1.0}))
+        .filter(l => l.a && l.b);
+
+    graphState.panX = canvas.width / 2;
+    graphState.panY = canvas.height / 2;
+    graphState.zoom = 1;
+    graphState.search = state.search || '';
+    graphState.selectedId = null;
+
+    _graphInitInteractions();
+    // Physics is expensive; pre-relax a little but keep it optional for big graphs.
+    const relaxIters = graphState.sim.length <= 700 ? 70 : 12;
+    for (let i = 0; i < relaxIters; i++) _graphStepPhysics(0.85);
+    _graphDraw();
+}
+
+function _graphStepPhysics(alpha = 1.0) {
+    const canvas = _graphCanvas();
+    const repulsion = (graphState.sim.length > 700) ? 120 : 180;
+    const damping = 0.86;
+    const target = 80;
+    const springK = 0.018;
+
+    const sim = graphState.sim;
+    if (sim.length > 700) {
+        // Fast approximate repulsion via spatial hashing (checks only local neighbor cells).
+        const cell = 120;
+        const grid = new Map();
+        const key = (cx, cy) => `${cx},${cy}`;
+        for (let i = 0; i < sim.length; i++) {
+            const n = sim[i];
+            const cx = Math.floor(n.x / cell);
+            const cy = Math.floor(n.y / cell);
+            const k = key(cx, cy);
+            if (!grid.has(k)) grid.set(k, []);
+            grid.get(k).push(i);
+        }
+        for (let i = 0; i < sim.length; i++) {
+            const a = sim[i];
+            const cx = Math.floor(a.x / cell);
+            const cy = Math.floor(a.y / cell);
+            for (let ox = -1; ox <= 1; ox++) {
+                for (let oy = -1; oy <= 1; oy++) {
+                    const idxs = grid.get(key(cx + ox, cy + oy));
+                    if (!idxs) continue;
+                    for (const j of idxs) {
+                        if (j === i) continue;
+                        const b = sim[j];
+                        const dx = a.x - b.x, dy = a.y - b.y;
+                        const d2 = dx*dx + dy*dy + 0.12;
+                        const f = (repulsion / d2) * alpha * 0.55;
+                        a.vx += dx*f; a.vy += dy*f;
+                    }
+                }
+            }
+        }
+    } else {
+        for (let i = 0; i < sim.length; i++) {
+            for (let j = i + 1; j < sim.length; j++) {
+                const a = sim[i], b = sim[j];
+                const dx = a.x - b.x, dy = a.y - b.y;
+                const d2 = dx*dx + dy*dy + 0.02;
+                const f = (repulsion / d2) * alpha;
+                a.vx += dx*f; a.vy += dy*f;
+                b.vx -= dx*f; b.vy -= dy*f;
+            }
+        }
+    }
+    for (const l of graphState.links) {
+        const a = l.a, b = l.b;
+        const dx = b.x - a.x, dy = b.y - a.y;
+        const dist = Math.sqrt(dx*dx + dy*dy) || 1;
+        const f = (dist - target) * springK * alpha;
+        const fx = (dx/dist) * f, fy = (dy/dist) * f;
+        a.vx += fx; a.vy += fy;
+        b.vx -= fx; b.vy -= fy;
+    }
+    for (const n of sim) {
+        n.vx *= damping; n.vy *= damping;
+        n.x += n.vx; n.y += n.vy;
+        const pad = 10;
+        const bx = canvas.width / graphState.zoom;
+        const by = canvas.height / graphState.zoom;
+        n.x = Math.max(-bx/2 + pad, Math.min(bx/2 - pad, n.x));
+        n.y = Math.max(-by/2 + pad, Math.min(by/2 - pad, n.y));
+    }
+}
+
+function _graphEdgeColor(kind) {
+    const k = (kind || '').toString().toLowerCase();
+    if (k.includes('tag')) return '#7c3aed';
+    if (k.includes('host')) return '#f59e0b';
+    if (k.includes('ref')) return '#10b981';
+    return '#3a3a55';
+}
+
+function _graphDraw() {
+    const canvas = _graphCanvas();
+    const ctx = _graphCtx();
+    const hint = document.getElementById('graphHint');
+
+    ctx.clearRect(0,0,canvas.width,canvas.height);
+    ctx.save();
+    ctx.translate(graphState.panX, graphState.panY);
+    ctx.scale(graphState.zoom, graphState.zoom);
+
+    for (const l of graphState.links) {
+        const term = (graphState.search || '').trim();
+        const isMatchEdge = term && (_graphMatches(l.a, term) || _graphMatches(l.b, term));
+        const w = Math.max(0.2, Math.min(3.0, (l.weight || 1.0)));
+        ctx.lineWidth = (0.6 + w) / graphState.zoom;
+        ctx.globalAlpha = isMatchEdge ? 0.85 : (0.25 + Math.min(0.35, w * 0.18));
+        ctx.strokeStyle = isMatchEdge ? '#f7d154' : _graphEdgeColor(l.kind);
+        ctx.beginPath();
+        ctx.moveTo(l.a.x, l.a.y);
+        ctx.lineTo(l.b.x, l.b.y);
+        ctx.stroke();
+    }
+    ctx.globalAlpha = 1.0;
+
+    const term = (graphState.search || '').trim();
+    let matches = 0;
+    for (const n of graphState.sim) {
+        const r = _graphNodeRadius(n);
+        const isMatch = _graphMatches(n, term);
+        if (isMatch) matches++;
+
+        if (isMatch) {
+            ctx.beginPath();
+            ctx.strokeStyle = '#f7d154';
+            ctx.lineWidth = 3 / graphState.zoom;
+            ctx.arc(n.x, n.y, r + 3, 0, Math.PI*2);
+            ctx.stroke();
+        }
+
+        if (graphState.selectedId === n.id) {
+            ctx.beginPath();
+            ctx.strokeStyle = '#ffffff';
+            ctx.lineWidth = 2 / graphState.zoom;
+            ctx.arc(n.x, n.y, r + 2, 0, Math.PI*2);
+            ctx.stroke();
+        }
+
+        ctx.beginPath();
+        ctx.fillStyle = _graphNodeFill(n);
+        ctx.arc(n.x, n.y, r, 0, Math.PI*2);
+        ctx.fill();
+
+        // status/recency/lock border
+        let stroke = null;
+        const v = (n.verdict || '').toString();
+        if (v === 'confirmed_true') stroke = '#bbf7d0';
+        else if (v === 'confirmed_false') stroke = '#fecaca';
+        else if (v) stroke = '#c7d2fe';
+
+        const now = Date.now();
+        const created = Date.parse(n.created || '') || 0;
+        const modified = Date.parse(n.modified || '') || 0;
+        const isNew = created && (now - created) < (30 * 60 * 1000);
+        const isHot = modified && (now - modified) < (10 * 60 * 1000);
+        if (isNew) stroke = '#f7d154';
+        if (isHot) stroke = '#ffffff';
+        if (n.predict_locked) stroke = '#a78bfa';
+
+        if (stroke) {
+            ctx.beginPath();
+            ctx.strokeStyle = stroke;
+            ctx.lineWidth = (n.predict_locked ? 3 : 1.5) / graphState.zoom;
+            ctx.arc(n.x, n.y, r + 0.8, 0, Math.PI*2);
+            ctx.stroke();
+        }
+    }
+
+    ctx.restore();
+    hint.textContent = `nodes=${graphState.nodes.length} edges=${graphState.edges.length}` + (term ? ` | match=${matches}` : '');
+}
+
+function _graphLoop() {
+    if (!graphState.physicsOn) return;
+    const speed = 0.25 + (Math.max(0, Math.min(100, graphState.physicsStrength || 60)) / 100) * 0.95;
+    _graphStepPhysics(speed);
+    _graphDraw();
+    graphState.raf = requestAnimationFrame(_graphLoop);
+}
+
+function setPhysicsStrength(v) {
+    const n = Math.max(0, Math.min(100, parseInt(v || '0', 10)));
+    graphState.physicsStrength = n;
+    localStorage.setItem('physicsStrength', String(n));
+    const el = document.getElementById('physicsStrengthVal');
+    if (el) el.textContent = String(n);
+}
+
+function toggleGraphPhysics() {
+    graphState.physicsOn = !graphState.physicsOn;
+    const b = document.getElementById('btnGraphPhysics');
+    const fast = (graphState.sim || []).length > 700;
+    b.textContent = `Physics: ${graphState.physicsOn ? ('on' + (fast ? ' (fast)' : '')) : 'off'}`;
+    b.classList.toggle('primary', graphState.physicsOn);
+    if (graphState.physicsOn) {
+        if (graphState.raf) cancelAnimationFrame(graphState.raf);
+        graphState.raf = requestAnimationFrame(_graphLoop);
+    } else {
+        if (graphState.raf) cancelAnimationFrame(graphState.raf);
+        graphState.raf = null;
+        _graphDraw();
+    }
+}
+
+function resetGraphView() {
+    const canvas = _graphCanvas();
+    graphState.panX = canvas.width / 2;
+    graphState.panY = canvas.height / 2;
+    graphState.zoom = 1;
+    _graphDraw();
+}
+
+function fitGraphView() {
+    const canvas = _graphCanvas();
+    if (!graphState.sim.length) return;
+    let minX = Infinity, minY = Infinity, maxX = -Infinity, maxY = -Infinity;
+    for (const n of graphState.sim) {
+        minX = Math.min(minX, n.x); minY = Math.min(minY, n.y);
+        maxX = Math.max(maxX, n.x); maxY = Math.max(maxY, n.y);
+    }
+    const w = Math.max(1, maxX - minX);
+    const h = Math.max(1, maxY - minY);
+    const zx = (canvas.width - 40) / w;
+    const zy = (canvas.height - 40) / h;
+    graphState.zoom = Math.max(0.35, Math.min(2.5, Math.min(zx, zy)));
+    graphState.panX = canvas.width / 2;
+    graphState.panY = canvas.height / 2;
+    _graphDraw();
+}
+
+function graphApplySearch(term) {
+    graphState.search = term || '';
+    _graphDraw();
+}
+
+// Real-time updates via SSE
+function startEvents() {
+    try {
+        const es = new EventSource('/api/events');
+        es.onmessage = (msg) => {
+            try {
+                state.lastEvent = JSON.parse(msg.data);
+                updateStatsFromEvent(state.lastEvent);
+                if (state.view === 'status') {
+                    // refresh status panels without heavy re-render: just rerun loadStatus occasionally
+                    loadStatus();
+                }
+                if (state.view === 'graph') {
+                    // fetch graph less often (every ~15s)
+                    const t = Date.now();
+                    if (!state._lastGraphFetch || (t - state._lastGraphFetch) > 15000) {
+                        state._lastGraphFetch = t;
+                        loadGraph();
+                    }
+                }
+            } catch (e) {}
+        };
+        es.onerror = () => {
+            // keep UI usable even if SSE drops
+        };
+    } catch (e) {}
+}
+
+startEvents();
+
 function renderCards() {
     const el = document.getElementById('cards');
     el.innerHTML = state.items.map(item => `
         <div class="card ${item.confirmed ? 'confirmed' : ''} ${item.rejections > 0 ? 'rejected' : ''}" data-id="${item.id}">
             <div class="card-header">
                 <span class="conf-badge" style="background:hsl(${item.confidence*120},70%,40%)">${Math.round(item.confidence*100)}%</span>
+                ${renderVerdictPill(item)}
                 <span class="tags">${item.tags.map(t => '<span class="tag">'+t+'</span>').join('')}</span>
                 <span class="date">${fmtDate(item.created)}</span>
             </div>
@@ -123,6 +976,19 @@ function renderCards() {
     `).join('');
 }
 
+function renderVerdictPill(item) {
+    const v = (item.verdict || '').toString();
+    if (!v) return '';
+    let cls = 'v-unknown';
+    let label = v;
+    if (v === 'confirmed_true') { cls = 'v-true'; label = 'TRUE'; }
+    else if (v === 'confirmed_false') { cls = 'v-false'; label = 'FALSE'; }
+    else if (v === 'probable_true') { cls = 'v-prob-true'; label = 'LIKELY'; }
+    else if (v === 'probable_false') { cls = 'v-prob-false'; label = 'UNLIKELY'; }
+    else if (v === 'unknown') { cls = 'v-unknown'; label = 'UNKNOWN'; }
+    return `<span class="verdict-pill ${cls}">${label}</span>`;
+}
+
 function fmtDate(iso) {
     const d = new Date(iso);
     return `${d.getDate().toString().padStart(2,'0')}.${(d.getMonth()+1).toString().padStart(2,'0')} ${d.getHours().toString().padStart(2,'0')}:${d.getMinutes().toString().padStart(2,'0')}`;
@@ -144,6 +1010,8 @@ async function confirm(id, ev) {
         body: new URLSearchParams({reason})
     });
     await loadCards(); await loadStats();
+    if (state.view === 'graph') loadGraph();
+    if (state.view === 'status') loadStatus();
 }
 
 async function reject(id, ev) {
@@ -155,6 +1023,8 @@ async function reject(id, ev) {
         body: new URLSearchParams({reason})
     });
     await loadCards(); await loadStats();
+    if (state.view === 'graph') loadGraph();
+    if (state.view === 'status') loadStatus();
 }
 
 async function refresh(id, ev) {
@@ -175,6 +1045,8 @@ async function createEngram() {
     document.getElementById('newContent').value = '';
     document.getElementById('newTags').value = '';
     await loadCards(); await loadStats();
+    if (state.view === 'graph') loadGraph();
+    if (state.view === 'status') loadStatus();
 }
 
 async function showDetail(id) {
@@ -220,6 +1092,7 @@ document.getElementById('searchInput').addEventListener('input', (e) => {
     state.search = e.target.value;
     state.offset = 0;
     loadCards();
+    if (state.view === 'graph') graphApplySearch(state.search);
 });
 
 document.getElementById('filterSelect').addEventListener('change', (e) => {