From 0484c6321a5b2ca08e544820f3a272010c683d5a Mon Sep 17 00:00:00 2001
From: Claudia <claudia.keller0001@gmail.com>
Date: Fri, 13 Feb 2026 11:52:25 +0100
Subject: [PATCH] feat(memory): add session memory persistence module

New cortex/memory/ module that provides:
- boot_assembler: builds BOOTSTRAP.md from threads, decisions, narrative
- thread_tracker: tracks conversation threads across sessions via NATS
- narrative_generator: daily narrative with Ollama LLM (fallback: structured)
- pre_compaction: snapshot pipeline before context compaction

CLI commands:
- cortex memory bootstrap [--dry-run] [--workspace DIR]
- cortex memory snapshot [--workspace DIR]
- cortex memory threads [--summary] [--hours N]

All paths configurable via WORKSPACE_DIR, NATS_URL, AGENT_NAME env vars.
No hardcoded paths. Works with any OpenClaw agent.

Fixes array/dict handling for empty threads.json and decisions.json.
---
 cortex/cli.py                        |  26 +++
 cortex/intelligence/loop.py          |   2 +-
 cortex/memory/__init__.py            |   1 +
 cortex/memory/boot_assembler.py      | 290 +++++++++++++++++++++++++++
 cortex/memory/common.py              | 122 +++++++++++
 cortex/memory/narrative_generator.py | 171 ++++++++++++++++
 cortex/memory/pre_compaction.py      | 110 ++++++++++
 cortex/memory/thread_tracker.py      | 267 ++++++++++++++++++++++++
 8 files changed, 988 insertions(+), 1 deletion(-)
 create mode 100644 cortex/memory/__init__.py
 create mode 100644 cortex/memory/boot_assembler.py
 create mode 100644 cortex/memory/common.py
 create mode 100644 cortex/memory/narrative_generator.py
 create mode 100644 cortex/memory/pre_compaction.py
 create mode 100644 cortex/memory/thread_tracker.py

diff --git a/cortex/cli.py b/cortex/cli.py
index f7cd196..f96dabd 100644
--- a/cortex/cli.py
+++ b/cortex/cli.py
@@ -22,6 +22,9 @@ Usage:
     cortex anomaly [--hours N] [--json]
     cortex predict [--learn] [--patterns]
     cortex monitor [--json]
+    cortex memory bootstrap [--dry-run] [--max-tokens N]
+    cortex memory snapshot [--dry-run]
+    cortex memory threads [--summary] [--hours N]
     cortex version
 """
 
@@ -121,6 +124,29 @@ def main():
         from cortex.monitor import main as monitor_main
         monitor_main()
 
+    elif cmd == "memory":
+        if len(sys.argv) < 2:
+            print("Usage: cortex memory bootstrap|snapshot|threads [options]")
+            sys.exit(1)
+        subcmd = sys.argv[1] if len(sys.argv) > 1 else ""
+        sys.argv = [f"cortex memory {subcmd}"] + sys.argv[2:]
+        if subcmd == "bootstrap":
+            from cortex.memory.boot_assembler import main as boot_main
+            boot_main()
+        elif subcmd == "snapshot":
+            from cortex.memory.pre_compaction import main as snap_main
+            snap_main()
+        elif subcmd == "threads":
+            # Default to --show mode for display
+            if "--hours" not in sys.argv and "--dry-run" not in sys.argv:
+                sys.argv.append("--show")
+            from cortex.memory.thread_tracker import main as thread_main
+            thread_main()
+        else:
+            print(f"Unknown memory subcommand: {subcmd}")
+            print("Available: bootstrap, snapshot, threads")
+            sys.exit(1)
+
     elif cmd in ("-h", "--help", "help"):
         print(__doc__.strip())
 
diff --git a/cortex/intelligence/loop.py b/cortex/intelligence/loop.py
index 98930af..b463667 100644
--- a/cortex/intelligence/loop.py
+++ b/cortex/intelligence/loop.py
@@ -309,7 +309,7 @@ def step_extract(state: LoopState, events: list) -> dict:
     spec.loader.exec_module(em)
 
     # Try LLM batch extraction first
-    from llm_extractor import extract_entities_llm_batch, is_available as llm_available
+    from cortex.llm_extractor import extract_entities_llm_batch, is_available as llm_available
     use_llm = os.environ.get("DARKPLEX_EXTRACTOR", "auto").lower() in ("llm", "auto")
     llm_ok = use_llm and llm_available()
     if llm_ok:
diff --git a/cortex/memory/__init__.py b/cortex/memory/__init__.py
new file mode 100644
index 0000000..268ed01
--- /dev/null
+++ b/cortex/memory/__init__.py
@@ -0,0 +1 @@
+"""Memory module — session memory persistence for OpenClaw agents."""
diff --git a/cortex/memory/boot_assembler.py b/cortex/memory/boot_assembler.py
new file mode 100644
index 0000000..fadb6ff
--- /dev/null
+++ b/cortex/memory/boot_assembler.py
@@ -0,0 +1,290 @@
+"""Boot Assembler — Query-driven boot context generator.
+
+Reads threads, decisions, narrative, and knowledge to assemble a dense BOOTSTRAP.md.
+All paths derived from WORKSPACE_DIR. No hardcoded paths.
+"""
+
+import json
+import os
+import subprocess
+from datetime import datetime, timezone, timedelta
+from pathlib import Path
+
+from .common import (
+    get_workspace_dir, get_reboot_dir, get_agent_name,
+    load_json, load_facts,
+)
+
+DEFAULT_MAX_CHARS = 16000  # ~4000 tokens
+
+
+def _load_threads_data(reboot_dir: Path) -> dict:
+    data = load_json(reboot_dir / "threads.json")
+    if isinstance(data, list):
+        return {"threads": data}
+    return data if isinstance(data, dict) else {}
+
+
+def _get_open_threads(reboot_dir: Path, limit: int = 7) -> list[dict]:
+    data = _load_threads_data(reboot_dir)
+    threads = [t for t in data.get("threads", []) if t.get("status") == "open"]
+    priority_order = {"critical": 0, "high": 1, "medium": 2, "low": 3}
+    threads.sort(key=lambda t: (
+        priority_order.get(t.get("priority", "low"), 3),
+        -(datetime.fromisoformat(
+            t.get("last_activity", "2000-01-01T00:00:00Z").replace("Z", "+00:00")
+        ).timestamp())
+    ))
+    return threads[:limit]
+
+
+def _integrity_warning(reboot_dir: Path) -> str:
+    data = _load_threads_data(reboot_dir)
+    integrity = data.get("integrity", {})
+    last_ts = integrity.get("last_nats_timestamp")
+    if not last_ts:
+        return "⚠️ No integrity data — thread tracker may not have run yet."
+    try:
+        if len(last_ts) <= 10:
+            last_dt = datetime.strptime(last_ts, "%Y-%m-%d").replace(tzinfo=timezone.utc)
+        else:
+            last_dt = datetime.fromisoformat(last_ts.replace("Z", "+00:00"))
+        age_min = (datetime.now(timezone.utc) - last_dt).total_seconds() / 60
+        if age_min > 480:
+            return f"🚨 STALE DATA: Thread data is {age_min/60:.0f}h old."
+        elif age_min > 120:
+            return f"⚠️ Data staleness: Thread data is {age_min/60:.0f}h old."
+        return ""
+    except Exception:
+        return "⚠️ Could not parse integrity timestamp."
+
+
+def _load_hot_snapshot(reboot_dir: Path) -> str:
+    f = reboot_dir / "hot-snapshot.md"
+    try:
+        if not f.exists():
+            return ""
+        mtime = datetime.fromtimestamp(f.stat().st_mtime, tz=timezone.utc)
+        if datetime.now(timezone.utc) - mtime > timedelta(hours=1):
+            return ""
+        return f.read_text().strip()[:1000]
+    except Exception:
+        return ""
+
+
+def _load_decisions(reboot_dir: Path) -> list[dict]:
+    data = load_json(reboot_dir / "decisions.json")
+    if isinstance(data, list):
+        data = {"decisions": data}
+    if not isinstance(data, dict):
+        return []
+    cutoff = (datetime.now(timezone.utc) - timedelta(days=14)).strftime("%Y-%m-%d")
+    recent = [d for d in data.get("decisions", []) if d.get("date", "") >= cutoff]
+    return recent[-10:]
+
+
+def _load_narrative(reboot_dir: Path) -> str:
+    f = reboot_dir / "narrative.md"
+    try:
+        if not f.exists():
+            return ""
+        mtime = datetime.fromtimestamp(f.stat().st_mtime, tz=timezone.utc)
+        if datetime.now(timezone.utc) - mtime > timedelta(hours=36):
+            return ""
+        return f.read_text().strip()[:2000]
+    except Exception:
+        return ""
+
+
+def _query_knowledge_for_thread(thread: dict, facts: list[dict]) -> list[str]:
+    """Score facts by keyword overlap with thread."""
+    results = []
+    query_terms = thread.get("title", "") + " " + thread.get("summary", "")
+    query_words = set(query_terms.lower().split())
+
+    scored = []
+    for fact in facts:
+        text = fact.get("text", "").lower()
+        priority = fact.get("priority", "normal")
+        boost = {"critical": 4, "high": 2, "normal": 1, "low": 0.5}.get(priority, 1)
+        overlap = len(query_words & set(text.split()))
+        if overlap > 0:
+            score = overlap * boost
+            created = fact.get("created", "")
+            if created and created[:10] >= (datetime.now(timezone.utc) - timedelta(days=7)).strftime("%Y-%m-%d"):
+                score *= 1.5
+            scored.append((score, fact))
+
+    scored.sort(key=lambda x: -x[0])
+    for score, fact in scored[:3]:
+        prio = fact.get("priority", "?")
+        conf = fact.get("confidence", 1.0)
+        text = fact.get("text", "")[:150]
+        results.append(f"  [{prio}|{conf:.0%}] {text}")
+
+    return results[:5]
+
+
+def _get_execution_mode() -> str:
+    hour = datetime.now().hour
+    if 6 <= hour < 12:
+        return "Morning — brief, directive, efficient"
+    elif 12 <= hour < 18:
+        return "Afternoon — execution mode"
+    elif 18 <= hour < 22:
+        return "Evening — strategic, philosophical possible"
+    return "Night — emergencies only"
+
+
+def assemble(workspace: Path = None, max_chars: int = DEFAULT_MAX_CHARS,
+             facts_file: Path = None, calendar_cmd: list[str] = None,
+             wellbeing_file: Path = None) -> str:
+    """Assemble BOOTSTRAP.md content.
+
+    Args:
+        workspace: Workspace directory (default: WORKSPACE_DIR or cwd)
+        max_chars: Character budget
+        facts_file: Path to facts.jsonl for knowledge queries
+        calendar_cmd: Command to run for calendar events (optional)
+        wellbeing_file: Path to wellbeing.json (optional)
+    """
+    ws = workspace or get_workspace_dir()
+    reboot_dir = get_reboot_dir(ws)
+    agent = get_agent_name()
+    now = datetime.now(timezone.utc)
+    local_now = datetime.now()
+    parts = []
+
+    parts.append(f"# Context Briefing")
+    parts.append(f"Agent: {agent} | Generated: {now.isoformat()[:19]}Z | Local: {local_now.strftime('%H:%M')}")
+    parts.append("")
+
+    # State
+    parts.append("## ⚡ State")
+    parts.append(f"Mode: {_get_execution_mode()}")
+
+    # Wellbeing (optional)
+    if wellbeing_file:
+        wb = load_json(wellbeing_file)
+        if wb:
+            parts.append(f"Wellbeing: {wb.get('status', '?')} ({wb.get('overall', 0):.0%}) trend:{wb.get('history_trend', '?')}")
+
+    # Session mood
+    td = _load_threads_data(reboot_dir)
+    mood = td.get("session_mood", "neutral")
+    if mood != "neutral":
+        emoji = {"frustrated": "😤", "excited": "🔥", "tense": "⚡", "productive": "🔧", "exploratory": "🔬"}.get(mood, "")
+        parts.append(f"Last session mood: {mood} {emoji}")
+
+    warning = _integrity_warning(reboot_dir)
+    if warning:
+        parts.append(f"\n{warning}")
+
+    # Calendar (optional external command)
+    if calendar_cmd:
+        try:
+            result = subprocess.run(calendar_cmd, capture_output=True, text=True, timeout=10)
+            if result.returncode == 0 and result.stdout.strip():
+                parts.append(f"\n### 📅 Today")
+                parts.append("\n".join(result.stdout.strip().split("\n")[:10]))
+        except Exception:
+            pass
+    parts.append("")
+
+    # Hot snapshot
+    hot = _load_hot_snapshot(reboot_dir)
+    if hot:
+        parts.append("## 🔥 Last Session Snapshot")
+        parts.append(hot)
+        parts.append("")
+
+    # Narrative
+    narrative = _load_narrative(reboot_dir)
+    if narrative:
+        parts.append("## 📖 Narrative (last 24h)")
+        parts.append(narrative)
+        parts.append("")
+
+    # Threads + knowledge
+    threads = _get_open_threads(reboot_dir)
+    all_facts = load_facts(facts_file) if facts_file and facts_file.exists() else []
+
+    if threads:
+        parts.append("## 🧵 Active Threads")
+        for t in threads:
+            prio_emoji = {"critical": "🔴", "high": "🟠", "medium": "🟡", "low": "🔵"}.get(t.get("priority"), "⚪")
+            mood_tag = f" [{t.get('mood', '')}]" if t.get("mood", "neutral") != "neutral" else ""
+            parts.append(f"\n### {prio_emoji} {t['title']}{mood_tag}")
+            parts.append(f"Priority: {t.get('priority', '?')} | Last: {t.get('last_activity', '?')[:16]}")
+            parts.append(f"Summary: {t.get('summary', 'no summary')}")
+            if t.get("waiting_for"):
+                parts.append(f"⏳ Waiting for: {t['waiting_for']}")
+            if t.get("decisions"):
+                parts.append(f"Decisions: {', '.join(t['decisions'])}")
+            if all_facts:
+                knowledge = _query_knowledge_for_thread(t, all_facts)
+                if knowledge:
+                    parts.append("Knowledge:")
+                    parts.extend(knowledge)
+        parts.append("")
+
+    # Decisions
+    decisions = _load_decisions(reboot_dir)
+    if decisions:
+        parts.append("## 🎯 Recent Decisions")
+        for d in decisions:
+            ie = {"critical": "🔴", "high": "🟠", "medium": "🟡"}.get(d.get("impact"), "⚪")
+            parts.append(f"- {ie} **{d['what']}** ({d.get('date', '?')})")
+            if d.get("why"):
+                parts.append(f"  Why: {d['why'][:100]}")
+        parts.append("")
+
+    # Footer
+    parts.append("---")
+    parts.append(f"_Boot context | {len(threads)} active threads | {len(decisions)} recent decisions_")
+
+    result = "\n".join(parts)
+    if len(result) > max_chars:
+        result = result[:max_chars] + "\n\n_[truncated to token budget]_"
+    return result
+
+
+def run(dry_run: bool = False, max_tokens: int = 4000, workspace: Path = None,
+        facts_file: Path = None, calendar_cmd: list[str] = None,
+        wellbeing_file: Path = None):
+    """Run boot assembler."""
+    ws = workspace or get_workspace_dir()
+    output_file = ws / "BOOTSTRAP.md"
+
+    bootstrap = assemble(
+        workspace=ws,
+        max_chars=max_tokens * 4,
+        facts_file=facts_file,
+        calendar_cmd=calendar_cmd,
+        wellbeing_file=wellbeing_file,
+    )
+
+    if dry_run:
+        print(bootstrap)
+        print(f"\n--- Stats: {len(bootstrap)} chars, ~{len(bootstrap)//4} tokens ---")
+    else:
+        output_file.write_text(bootstrap)
+        print(f"✅ BOOTSTRAP.md written ({len(bootstrap)} chars, ~{len(bootstrap)//4} tokens)")
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="Boot Assembler — Query-driven boot context")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--max-tokens", type=int, default=4000)
+    parser.add_argument("--workspace", type=str, help="Workspace directory")
+    parser.add_argument("--facts-file", type=str, help="Path to facts.jsonl")
+    args = parser.parse_args()
+
+    ws = Path(args.workspace) if args.workspace else None
+    ff = Path(args.facts_file) if args.facts_file else None
+    run(dry_run=args.dry_run, max_tokens=args.max_tokens, workspace=ws, facts_file=ff)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/cortex/memory/common.py b/cortex/memory/common.py
new file mode 100644
index 0000000..cb39d18
--- /dev/null
+++ b/cortex/memory/common.py
@@ -0,0 +1,122 @@
+"""Shared utilities for the memory module — NATS access, path helpers, JSON loading."""
+
+import json
+import logging
+import os
+from datetime import datetime, timezone
+from pathlib import Path
+
+
+def get_workspace_dir() -> Path:
+    """Get workspace directory from WORKSPACE_DIR env or cwd."""
+    return Path(os.environ.get("WORKSPACE_DIR", os.getcwd()))
+
+
+def get_agent_name() -> str:
+    """Get agent name from AGENT_NAME env or 'agent'."""
+    return os.environ.get("AGENT_NAME", "agent")
+
+
+def get_reboot_dir(workspace: Path = None) -> Path:
+    """Get memory/reboot directory, creating if needed."""
+    ws = workspace or get_workspace_dir()
+    d = ws / "memory" / "reboot"
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def get_nats_credentials(workspace: Path = None) -> dict:
+    """Load NATS credentials from env vars or config file.
+
+    Priority: env vars > config file at WORKSPACE_DIR/config/nats/credentials.env
+    Returns dict with keys: url, user, password
+    """
+    url = os.environ.get("NATS_URL", "")
+    user = os.environ.get("NATS_USER", "")
+    password = os.environ.get("NATS_PASSWORD", "")
+
+    if not (url and user and password):
+        ws = workspace or get_workspace_dir()
+        creds_file = ws / "config" / "nats" / "credentials.env"
+        if creds_file.exists():
+            for line in creds_file.read_text().strip().split("\n"):
+                if "=" in line and not line.startswith("#"):
+                    k, v = line.split("=", 1)
+                    k, v = k.strip(), v.strip().strip('"')
+                    if k == "NATS_URL" and not url:
+                        url = v
+                    elif k in ("NATS_USER",) and not user:
+                        user = v
+                    elif k in ("NATS_PASSWORD", "NATS_CLAUDIA_PW") and not password:
+                        password = v
+
+    return {
+        "url": url or "nats://localhost:4222",
+        "user": user,
+        "password": password,
+    }
+
+
+def get_nats_env(workspace: Path = None) -> dict:
+    """Return os.environ copy with NATS credentials set for nats CLI."""
+    creds = get_nats_credentials(workspace)
+    env = os.environ.copy()
+    if creds["user"]:
+        env["NATS_USER"] = creds["user"]
+    if creds["password"]:
+        env["NATS_PASSWORD"] = creds["password"]
+    if creds["url"]:
+        env["NATS_URL"] = creds["url"]
+    return env
+
+
+def load_json(path: Path) -> dict:
+    """Load JSON file, returning empty dict on failure."""
+    try:
+        return json.loads(path.read_text())
+    except (FileNotFoundError, json.JSONDecodeError):
+        return {}
+
+
+def save_json(path: Path, data: dict):
+    """Atomically write JSON to file."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_suffix(".tmp")
+    tmp.write_text(json.dumps(data, indent=2, ensure_ascii=False))
+    tmp.rename(path)
+
+
+def load_facts(path: Path) -> list[dict]:
+    """Load facts from a JSONL file."""
+    if not path.exists():
+        return []
+    facts = []
+    for line in path.read_text().strip().split("\n"):
+        if not line.strip():
+            continue
+        try:
+            fact = json.loads(line)
+            if "text" not in fact and "fact" in fact:
+                fact["text"] = fact["fact"]
+            facts.append(fact)
+        except json.JSONDecodeError:
+            continue
+    return facts
+
+
+def setup_logging(name: str, workspace: Path = None) -> logging.Logger:
+    """Configure logging to workspace/logs/ and stderr."""
+    ws = workspace or get_workspace_dir()
+    log_dir = ws / "logs"
+    log_dir.mkdir(parents=True, exist_ok=True)
+    log_file = log_dir / f"{name}.log"
+
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s [%(levelname)s] %(message)s",
+        handlers=[
+            logging.FileHandler(log_file),
+            logging.StreamHandler(),
+        ],
+    )
+    return logging.getLogger(name)
diff --git a/cortex/memory/narrative_generator.py b/cortex/memory/narrative_generator.py
new file mode 100644
index 0000000..0db521f
--- /dev/null
+++ b/cortex/memory/narrative_generator.py
@@ -0,0 +1,171 @@
+"""Narrative Generator — Creates a human-readable story from recent activity.
+
+Reads daily notes and thread/decision data, compresses into a narrative.
+Optionally uses a local LLM (Ollama) for richer narratives with structured fallback.
+"""
+
+import json
+import os
+import subprocess
+import sys
+from datetime import datetime, timezone, timedelta
+from pathlib import Path
+
+from .common import get_workspace_dir, get_reboot_dir, get_agent_name, load_json
+
+# Configurable via env
+OLLAMA_MODEL = os.environ.get("OLLAMA_MODEL", "mistral:7b")
+OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434")
+
+
+def load_daily_notes(workspace: Path, hours: int = 24) -> str:
+    """Load relevant daily notes."""
+    parts = []
+    today = datetime.now().strftime("%Y-%m-%d")
+    yesterday = (datetime.now() - timedelta(days=1)).strftime("%Y-%m-%d")
+
+    for date in [yesterday, today]:
+        note_file = workspace / "memory" / f"{date}.md"
+        if note_file.exists():
+            parts.append(f"## {date}\n{note_file.read_text()[:4000]}")
+
+    return "\n\n".join(parts)
+
+
+def load_threads(workspace: Path) -> list[dict]:
+    """Load threads for context."""
+    data = load_json(get_reboot_dir(workspace) / "threads.json")
+    return data.get("threads", [])
+
+
+def load_decisions(workspace: Path) -> list[dict]:
+    """Load recent decisions."""
+    data = load_json(get_reboot_dir(workspace) / "decisions.json")
+    today = datetime.now().strftime("%Y-%m-%d")
+    yesterday = (datetime.now() - timedelta(days=1)).strftime("%Y-%m-%d")
+    return [d for d in data.get("decisions", []) if d.get("date", "") >= yesterday]
+
+
+def generate_structured(notes: str, threads: list[dict], decisions: list[dict]) -> str:
+    """Generate narrative without LLM — structured approach."""
+    now = datetime.now()
+    parts = [f"*{now.strftime('%A, %d. %B %Y')} — Narrative*\n"]
+
+    open_threads = [t for t in threads if t.get("status") == "open"]
+    closed_threads = [t for t in threads if t.get("status") == "closed"
+                      and t.get("last_activity", "")[:10] >= (now - timedelta(days=1)).strftime("%Y-%m-%d")]
+
+    if closed_threads:
+        parts.append("**Completed:**")
+        for t in closed_threads:
+            parts.append(f"- ✅ {t['title']}: {t.get('summary', '')[:100]}")
+        parts.append("")
+
+    if open_threads:
+        parts.append("**Open:**")
+        for t in open_threads:
+            prio = t.get("priority", "?")
+            parts.append(f"- {'🔴' if prio == 'critical' else '🟡'} {t['title']}: {t.get('summary', '')[:150]}")
+            if t.get("waiting_for"):
+                parts.append(f"  ⏳ {t['waiting_for']}")
+        parts.append("")
+
+    if decisions:
+        parts.append("**Decisions:**")
+        for d in decisions:
+            parts.append(f"- {d['what']} — {d.get('why', '')[:80]}")
+        parts.append("")
+
+    if notes:
+        parts.append("**Timeline:**")
+        for line in notes.split("\n"):
+            line = line.strip()
+            if line.startswith("## ") and not line.startswith("## 20"):
+                parts.append(f"- {line[3:]}")
+            elif line.startswith("### "):
+                parts.append(f"  - {line[4:]}")
+        parts.append("")
+
+    return "\n".join(parts)
+
+
+def generate_llm(notes: str, threads: list[dict], decisions: list[dict],
+                  model: str = None) -> str | None:
+    """Generate narrative using local Ollama LLM. Returns None on failure."""
+    agent = get_agent_name()
+    model = model or OLLAMA_MODEL
+
+    context = f"""You are a memory system for an AI agent named {agent}. Write a brief, dense narrative (max 500 words) of what happened in the last 24 hours. Be specific — names, decisions, outcomes. No fluff.
+
+Daily Notes:
+{notes[:3000]}
+
+Open Threads:
+{json.dumps([{"title": t["title"], "summary": t.get("summary", ""), "priority": t.get("priority")} for t in threads if t.get("status") == "open"], indent=2)[:1000]}
+
+Decisions:
+{json.dumps(decisions[:5], indent=2)[:500]}
+
+Write the narrative now. Be concise."""
+
+    try:
+        result = subprocess.run(
+            ["ollama", "run", model, context],
+            capture_output=True, text=True, timeout=60
+        )
+        if result.returncode == 0 and len(result.stdout.strip()) > 50:
+            return result.stdout.strip()[:2000]
+    except Exception:
+        pass
+
+    return None
+
+
+def run(hours: int = 24, dry_run: bool = False, no_llm: bool = False,
+        model: str = None, workspace: Path = None) -> str:
+    """Run narrative generator. Returns the narrative text."""
+    ws = workspace or get_workspace_dir()
+    reboot_dir = get_reboot_dir(ws)
+    narrative_file = reboot_dir / "narrative.md"
+
+    print("📖 Narrative Generator — building story...")
+
+    notes = load_daily_notes(ws, hours)
+    threads = load_threads(ws)
+    decisions = load_decisions(ws)
+
+    narrative = None
+    if not no_llm:
+        narrative = generate_llm(notes, threads, decisions, model=model)
+        if narrative:
+            print("  Used LLM for narrative")
+
+    if not narrative:
+        narrative = generate_structured(notes, threads, decisions)
+        print("  Used structured narrative (no LLM)")
+
+    if dry_run:
+        print("\n" + narrative)
+    else:
+        narrative_file.write_text(narrative)
+        print(f"  ✅ narrative.md written ({len(narrative)} chars)")
+
+    return narrative
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="Narrative Generator")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--hours", type=int, default=24)
+    parser.add_argument("--no-llm", action="store_true", help="Skip LLM, use structured only")
+    parser.add_argument("--model", type=str, default=None, help=f"Ollama model (default: {OLLAMA_MODEL})")
+    parser.add_argument("--workspace", type=str, help="Workspace directory")
+    args = parser.parse_args()
+
+    ws = Path(args.workspace) if args.workspace else None
+    run(hours=args.hours, dry_run=args.dry_run, no_llm=args.no_llm, model=args.model, workspace=ws)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/cortex/memory/pre_compaction.py b/cortex/memory/pre_compaction.py
new file mode 100644
index 0000000..b4b4e4f
--- /dev/null
+++ b/cortex/memory/pre_compaction.py
@@ -0,0 +1,110 @@
+"""Pre-Compaction Snapshot — Captures the "hot zone" before memory loss.
+
+Orchestrates: thread tracker → hot snapshot → narrative → boot assembler.
+"""
+
+import json
+import subprocess
+from datetime import datetime, timezone
+from pathlib import Path
+
+from .common import get_workspace_dir, get_reboot_dir, get_nats_env
+from .thread_tracker import run as run_thread_tracker
+from .narrative_generator import run as run_narrative
+from .boot_assembler import run as run_boot_assembler
+
+
+def fetch_recent_messages(count: int = 20, workspace: Path = None) -> list[str]:
+    """Fetch last N messages from NATS for snapshot."""
+    ws = workspace or get_workspace_dir()
+    env = get_nats_env(ws)
+    messages = []
+
+    try:
+        result = subprocess.run(
+            ["nats", "stream", "get", "openclaw-events", "--last", str(count), "--raw"],
+            capture_output=True, text=True, timeout=15, env=env
+        )
+        if result.returncode == 0 and result.stdout.strip():
+            for line in result.stdout.strip().split("\n"):
+                try:
+                    evt = json.loads(line)
+                    content = evt.get("content", evt.get("message", evt.get("text", "")))
+                    sender = evt.get("sender", evt.get("agent", evt.get("role", "?")))
+                    if content and len(content.strip()) > 3:
+                        short = content.strip()[:200]
+                        if len(content.strip()) > 200:
+                            short += "..."
+                        messages.append(f"[{sender}] {short}")
+                except json.JSONDecodeError:
+                    continue
+    except Exception as e:
+        messages.append(f"(NATS fetch failed: {e})")
+
+    return messages
+
+
+def build_snapshot(messages: list[str]) -> str:
+    """Build the hot snapshot markdown."""
+    now = datetime.now(timezone.utc)
+    parts = [
+        f"# Hot Snapshot — {now.isoformat()[:19]}Z",
+        "## Last ~30min before compaction",
+        "",
+    ]
+    if messages:
+        parts.append("**Recent conversation:**")
+        for msg in messages[-15:]:
+            parts.append(f"- {msg}")
+    else:
+        parts.append("(No recent messages captured)")
+    parts.append("")
+    return "\n".join(parts)
+
+
+def run(dry_run: bool = False, workspace: Path = None, **assembler_kwargs):
+    """Run the full pre-compaction pipeline."""
+    ws = workspace or get_workspace_dir()
+    reboot_dir = get_reboot_dir(ws)
+    hot_snapshot_file = reboot_dir / "hot-snapshot.md"
+
+    print("🔥 Pre-Compaction Snapshot — capturing hot zone...")
+
+    # 1. Thread tracker (last 1h)
+    print("  1/4 Thread tracker (last 1h)...")
+    run_thread_tracker(hours=1, workspace=ws)
+
+    # 2. Hot snapshot
+    print("  2/4 Capturing recent messages...")
+    messages = fetch_recent_messages(20, ws)
+    snapshot = build_snapshot(messages)
+    if dry_run:
+        print(snapshot)
+    else:
+        hot_snapshot_file.write_text(snapshot)
+        print(f"  ✅ hot-snapshot.md written ({len(snapshot)} chars)")
+
+    # 3. Narrative (no LLM for speed during compaction)
+    print("  3/4 Narrative generator...")
+    run_narrative(no_llm=True, workspace=ws)
+
+    # 4. Boot assembler
+    print("  4/4 Boot assembler...")
+    run_boot_assembler(workspace=ws, **assembler_kwargs)
+
+    print("🔥 Pre-Compaction Snapshot — done!")
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="Pre-Compaction Snapshot")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--workspace", type=str, help="Workspace directory")
+    args = parser.parse_args()
+
+    ws = Path(args.workspace) if args.workspace else None
+    run(dry_run=args.dry_run, workspace=ws)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/cortex/memory/thread_tracker.py b/cortex/memory/thread_tracker.py
new file mode 100644
index 0000000..c2a6ba6
--- /dev/null
+++ b/cortex/memory/thread_tracker.py
@@ -0,0 +1,267 @@
+"""Thread Tracker — Tracks open conversation threads from NATS events.
+
+Reads recent NATS events and extracts/updates conversation threads.
+Thread detection heuristics: topic shifts, decision markers, status markers, waiting markers.
+"""
+
+import json
+import re
+import subprocess
+import sys
+from datetime import datetime, timezone, timedelta
+from pathlib import Path
+
+from .common import get_workspace_dir, get_reboot_dir, get_nats_env, load_json, save_json
+
+# Patterns for thread detection
+DECISION_PATTERNS = [
+    r"(?:decided|entschieden|decision|beschlossen|agreed|let'?s do|machen wir|wir machen)",
+    r"(?:the plan is|der plan ist|approach:|ansatz:)",
+]
+CLOSE_PATTERNS = [
+    r"(?:done|erledigt|fixed|gefixt|gelöst|solved|closed|fertig|works|funktioniert|✅)",
+]
+WAIT_PATTERNS = [
+    r"(?:waiting for|warte auf|blocked by|blockiert durch|need.*first|brauche.*erst)",
+]
+TOPIC_PATTERNS = [
+    r"(?:back to|zurück zu|jetzt zu|now about|regarding|bzgl\.?|wegen)\s+(\w[\w\s-]{2,30})",
+]
+
+MOOD_PATTERNS = {
+    "frustrated": r"(?:fuck|shit|mist|nervig|genervt|damn|wtf|argh|schon wieder|zum kotzen|sucks)",
+    "excited": r"(?:geil|nice|awesome|krass|boom|läuft|yes!|🎯|🚀|perfekt|brilliant|mega|sick)",
+    "tense": r"(?:vorsicht|careful|risky|heikel|kritisch|dringend|urgent|achtung|gefährlich)",
+    "productive": r"(?:erledigt|done|fixed|works|fertig|deployed|✅|gebaut|shipped|läuft)",
+    "exploratory": r"(?:was wäre wenn|what if|könnte man|idea|idee|maybe|vielleicht|experiment)",
+}
+
+
+def detect_mood(text: str) -> str:
+    """Detect overall mood from text. Last match wins."""
+    if not text:
+        return "neutral"
+    text_lower = text.lower()
+    last_mood = "neutral"
+    last_pos = -1
+    for mood, pattern in MOOD_PATTERNS.items():
+        for m in re.finditer(pattern, text_lower):
+            if m.start() > last_pos:
+                last_pos = m.start()
+                last_mood = mood
+    return last_mood
+
+
+def fetch_recent_events(hours: int = 4, workspace: Path = None) -> list[dict]:
+    """Fetch recent user/agent messages from NATS, with daily note fallback."""
+    ws = workspace or get_workspace_dir()
+    env = get_nats_env(ws)
+    events = []
+
+    try:
+        result = subprocess.run(
+            ["nats", "stream", "get", "openclaw-events", "--last", "200", "--raw"],
+            capture_output=True, text=True, timeout=15, env=env
+        )
+        if result.returncode == 0 and result.stdout.strip():
+            for line in result.stdout.strip().split("\n"):
+                try:
+                    events.append(json.loads(line))
+                except json.JSONDecodeError:
+                    continue
+    except Exception:
+        pass
+
+    # Fallback: read from daily notes
+    if not events:
+        today = datetime.now().strftime("%Y-%m-%d")
+        yesterday = (datetime.now() - timedelta(days=1)).strftime("%Y-%m-%d")
+        for date in [today, yesterday]:
+            note_file = ws / "memory" / f"{date}.md"
+            if note_file.exists():
+                events.append({
+                    "type": "daily_note",
+                    "date": date,
+                    "content": note_file.read_text()[:5000]
+                })
+
+    return events
+
+
+def extract_thread_signals(text: str) -> dict:
+    """Extract thread-related signals from text."""
+    text_lower = text.lower()
+    signals = {"decisions": [], "closures": [], "waits": [], "topics": []}
+
+    for pattern in DECISION_PATTERNS:
+        for m in re.finditer(pattern, text_lower):
+            start = max(0, m.start() - 50)
+            end = min(len(text), m.end() + 100)
+            signals["decisions"].append(text[start:end].strip())
+
+    for pattern in CLOSE_PATTERNS:
+        if re.search(pattern, text_lower):
+            signals["closures"].append(True)
+
+    for pattern in WAIT_PATTERNS:
+        for m in re.finditer(pattern, text_lower):
+            start = m.start()
+            end = min(len(text), m.end() + 80)
+            signals["waits"].append(text[start:end].strip())
+
+    for pattern in TOPIC_PATTERNS:
+        for m in re.finditer(pattern, text_lower):
+            if m.group(1):
+                signals["topics"].append(m.group(1).strip())
+
+    return signals
+
+
+def process_events(events: list[dict], existing_threads: list[dict]) -> list[dict]:
+    """Process events and update thread list."""
+    threads = existing_threads.copy()
+
+    for evt in events:
+        content = ""
+        if isinstance(evt, dict):
+            content = evt.get("content", evt.get("message", evt.get("text", "")))
+            if not content and evt.get("type") == "daily_note":
+                content = evt.get("content", "")
+        if not content:
+            continue
+
+        signals = extract_thread_signals(content)
+
+        if signals["closures"]:
+            for t in threads:
+                if t["status"] == "open":
+                    thread_words = set(t.get("title", "").lower().split())
+                    content_words = set(content.lower().split())
+                    if len(thread_words & content_words) >= 2:
+                        t["status"] = "closed"
+                        t["last_activity"] = datetime.now(timezone.utc).isoformat()
+
+        for decision_ctx in signals["decisions"]:
+            for t in threads:
+                if t["status"] == "open":
+                    thread_words = set(t.get("title", "").lower().split())
+                    decision_words = set(decision_ctx.lower().split())
+                    if len(thread_words & decision_words) >= 2:
+                        if decision_ctx not in t.get("decisions", []):
+                            t.setdefault("decisions", []).append(decision_ctx[:100])
+
+    return threads
+
+
+def run(hours: int = 4, dry_run: bool = False, workspace: Path = None) -> dict:
+    """Run thread tracker. Returns the threads data dict."""
+    ws = workspace or get_workspace_dir()
+    reboot_dir = get_reboot_dir(ws)
+    threads_file = reboot_dir / "threads.json"
+
+    print(f"🧵 Thread Tracker — scanning last {hours}h of events...")
+
+    data = load_json(threads_file)
+    existing_threads = data.get("threads", [])
+
+    events = fetch_recent_events(hours, ws)
+    print(f"  Found {len(events)} events")
+
+    updated_threads = process_events(events, existing_threads)
+
+    # Prune closed threads older than 7 days
+    cutoff = (datetime.now(timezone.utc) - timedelta(days=7)).isoformat()
+    pruned = [t for t in updated_threads
+              if not (t.get("status") == "closed" and t.get("last_activity", "") < cutoff)]
+
+    data["threads"] = pruned
+    data["updated"] = datetime.now(timezone.utc).isoformat()
+    data["version"] = 2
+
+    # Integrity tracking
+    last_nats_ts = None
+    last_nats_seq = None
+    for evt in reversed(events):
+        if isinstance(evt, dict):
+            if evt.get("seq"):
+                last_nats_seq = evt["seq"]
+            ts = evt.get("timestamp", evt.get("time", evt.get("date")))
+            if ts:
+                last_nats_ts = ts
+                break
+    data["integrity"] = {
+        "last_nats_seq": last_nats_seq,
+        "last_nats_timestamp": last_nats_ts or datetime.now(timezone.utc).isoformat(),
+        "events_processed": len(events),
+        "source": "streaming" if events and not any(e.get("type") == "daily_note" for e in events) else "daily_notes",
+    }
+
+    all_content = " ".join(
+        evt.get("content", evt.get("message", evt.get("text", "")))
+        for evt in events if isinstance(evt, dict)
+    )
+    data["session_mood"] = detect_mood(all_content)
+
+    if dry_run:
+        print(json.dumps(data, indent=2, ensure_ascii=False))
+    else:
+        save_json(threads_file, data)
+        open_count = sum(1 for t in pruned if t.get("status") == "open")
+        print(f"  ✅ {len(pruned)} threads ({open_count} open), written to {threads_file}")
+
+    return data
+
+
+def show_threads(summary: bool = False, workspace: Path = None):
+    """Display current thread state."""
+    ws = workspace or get_workspace_dir()
+    threads_file = get_reboot_dir(ws) / "threads.json"
+    data = load_json(threads_file)
+    threads = data.get("threads", [])
+
+    if not threads:
+        print("No threads tracked.")
+        return
+
+    open_threads = [t for t in threads if t.get("status") == "open"]
+    closed_threads = [t for t in threads if t.get("status") == "closed"]
+
+    if summary:
+        print(f"Threads: {len(open_threads)} open, {len(closed_threads)} closed")
+        print(f"Session mood: {data.get('session_mood', 'neutral')}")
+        print(f"Last updated: {data.get('updated', '?')}")
+        return
+
+    prio_emoji = {"critical": "🔴", "high": "🟠", "medium": "🟡", "low": "🔵"}
+    for t in open_threads:
+        emoji = prio_emoji.get(t.get("priority"), "⚪")
+        print(f"{emoji} {t['title']} [{t.get('priority', '?')}]")
+        print(f"  {t.get('summary', '')[:120]}")
+        if t.get("waiting_for"):
+            print(f"  ⏳ {t['waiting_for']}")
+        print()
+
+    if closed_threads:
+        print(f"--- {len(closed_threads)} closed threads ---")
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="Thread Tracker")
+    parser.add_argument("--hours", type=int, default=4, help="Look back N hours")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--summary", action="store_true", help="Show summary only")
+    parser.add_argument("--show", action="store_true", help="Show threads without updating")
+    parser.add_argument("--workspace", type=str, help="Workspace directory")
+    args = parser.parse_args()
+
+    ws = Path(args.workspace) if args.workspace else None
+
+    if args.show or args.summary:
+        show_threads(summary=args.summary, workspace=ws)
+    else:
+        run(hours=args.hours, dry_run=args.dry_run, workspace=ws)
+
+
+if __name__ == "__main__":
+    main()