release v1.3: fix crash loop (FileNotFoundError), add content chunking

ERROR FIXED:
- FileNotFoundError on deleted session files caused 2,551 restarts/24h
- Embedding token overflow for long messages (exceeded 4K token limit)

ROOT CAUSES:
1. File opened before existence check - crash when OpenClaw deletes session files
2. No content truncation for embedding - long messages failed silently

FIXES:
1. Added pre-check for file existence before open()
2. Added try/catch for FileNotFoundError
3. Added chunk_text() for long content (6000 char chunks with 200 char overlap)
4. Each chunk stored separately with metadata (chunk_index, total_chunks)

IMPACT:
- Eliminates crash loop
- No memory loss for long messages
- Graceful recovery on session rotation

CHANGELOG.md

@@ -0,0 +1,97 @@
+# Changelog - openclaw-true-recall-base
+
+All notable changes to this project will be documented in this file.
+
+## [v1.3] - 2026-03-10
+
+### Fixed
+
+#### Critical: Crash Loop on Deleted Session Files
+
+**Error:** `FileNotFoundError: [Errno 2] No such file or directory: '/root/.openclaw/agents/main/sessions/daccff90-f889-44fa-ba8b-c8d7397e5241.jsonl'`
+
+**Root Cause:**
+- OpenClaw deletes session `.jsonl` files when `/new` or `/reset` is called
+- The watcher opened the file before checking existence
+- Between file detection and opening, the file was deleted
+- This caused unhandled `FileNotFoundError` → crash → systemd restart
+
+**Impact:** 2,551 restarts in 24 hours
+
+**Original Code (v1.2):**
+```python
+# Track file handle for re-opening
+f = open(session_file, 'r')  # CRASH HERE if file deleted
+f.seek(last_position)
+
+try:
+    while running:
+        if not session_file.exists():  # Check happens AFTER crash
+            ...
+```
+
+**Fix (v1.3):**
+```python
+# Check file exists before opening (handles deleted sessions)
+if not session_file.exists():
+    print(f"Session file gone: {session_file.name}, looking for new session...", file=sys.stderr)
+    return None
+
+# Track file handle for re-opening
+try:
+    f = open(session_file, 'r')
+    f.seek(last_position)
+except FileNotFoundError:
+    print(f"Session file removed during open: {session_file.name}", file=sys.stderr)
+    return None
+```
+
+#### Embedding Token Overflow
+
+**Error:** `Ollama API error 400: {"StatusCode":400,"Status":"400 Bad Request","error":"prompt too long; exceeded max context length by 4 tokens"}`
+
+**Root Cause:**
+- The embedding model `snowflake-arctic-embed2` has a 4,096 token limit (~16K chars)
+- Long messages were sent to embedding without truncation
+- The watcher's `get_embedding()` call passed full `turn['content']`
+
+**Impact:** Failed embedding generation, memory loss for long messages
+
+**Fix:**
+- Added `chunk_text()` function to split long content into 6,000 char overlapping chunks
+- Each chunk gets its own Qdrant point with `chunk_index` and `total_chunks` metadata
+- Overlap (200 chars) ensures search continuity
+- No data loss - all content stored
+
+### Changed
+
+- `store_to_qdrant()` now handles multiple chunks per turn
+- Each chunk stored with metadata: `chunk_index`, `total_chunks`, `full_content_length`
+
+---
+
+## [v1.2] - 2026-02-26
+
+### Fixed
+
+- Session rotation bug - added inactivity detection (30s threshold)
+- Improved file scoring to properly detect new sessions on `/new` or `/reset`
+
+---
+
+## [v1.1] - 2026-02-25
+
+### Added
+
+- 1-second mtime polling for session rotation
+
+---
+
+## [v1.0] - 2026-02-24
+
+### Added
+
+- Initial release
+- Real-time monitoring of OpenClaw sessions
+- Automatic embedding via local Ollama (snowflake-arctic-embed2)
+- Storage to Qdrant `memories_tr` collection

README.md

@@ -61,7 +61,7 @@ true-recall-base (REQUIRED)
     │   ├── Curator extracts gems → gems_tr
     │   └── Plugin injects gems into prompts
     │
-    └──▶ true-recall-blocks (ADDON)
+    └──▶ openclaw-true-recall-blocks (ADDON)
         ├── Topic clustering → topic_blocks_tr
         └── Contextual block retrieval
 
@@ -654,3 +654,108 @@ curl -s "http://10.0.0.40:6333/collections/memories_tr/points/scroll" \
 ---
 
 **Prerequisite for:** TrueRecall Gems, TrueRecall Blocks
+
+
+---
+
+## Upgrading from Older Versions
+
+This section covers full upgrades from older TrueRecall Base installations to the current version.
+
+### Version History
+
+| Version | Key Changes |
+|---------|-------------|
+| **v1.0** | Initial release - basic watcher |
+| **v1.1** | Session detection improvements |
+| **v1.2** | Priority-based session detection, lock file validation, backfill script |
+| **v1.3** | Offset persistence (resumes from last position), fixes duplicate processing |
+| **v1.4** | Current version - Memory backfill fix (Qdrant ids field), improved error handling |
+
+### Upgrade Paths
+
+#### From v1.0/v1.1/v1.2 → v1.4 (Current)
+
+If you have an older installation, follow these steps:
+
+```bash
+# Step 1: Backup existing configuration
+cp /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py    /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py.bak.$(date +%Y%m%d)
+
+cp /root/.openclaw/workspace/skills/qdrant-memory/scripts/config.json    /root/.openclaw/workspace/skills/qdrant-memory/scripts/config.json.bak.$(date +%Y%m%d)
+```
+
+```bash
+# Step 2: Stop the watcher
+pkill -f realtime_qdrant_watcher
+# Verify stopped
+ps aux | grep realtime_qdrant_watcher
+```
+
+```bash
+# Step 3: Download latest files (choose one source)
+
+# Option A: From GitLab (recommended)
+curl -o /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py   https://gitlab.com/mdkrush/openclaw-true-recall-base/-/raw/master/watcher/realtime_qdrant_watcher.py
+
+# Option B: From Gitea
+curl -o /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py   http://10.0.0.61:3000/SpeedyFoxAi/openclaw-true-recall-base/raw/branch/master/watcher/realtime_qdrant_watcher.py
+
+# Option C: From local clone (if you cloned the repo)
+cp /path/to/openclaw-true-recall-base/watcher/realtime_qdrant_watcher.py    /root/.openclaw/workspace/skills/qdrant-memory/scripts/
+```
+
+```bash
+# Step 4: Start the watcher
+python3 /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py --daemon
+```
+
+```bash
+# Step 5: Verify installation
+ps aux | grep realtime_qdrant_watcher
+curl -s "http://10.0.0.40:6333/collections/memories_tr/points/scroll"   -H "Content-Type: application/json" -d '{"limit": 3}' | jq '.result.points[0].payload.timestamp'
+```
+
+### Upgrading with Git (If You Cloned the Repository)
+
+```bash
+# Navigate to your clone
+cd /path/to/openclaw-true-recall-base
+git pull origin master
+
+# Stop current watcher
+pkill -f realtime_qdrant_watcher
+
+# Copy updated files to OpenClaw
+cp watcher/realtime_qdrant_watcher.py /root/.openclaw/workspace/skills/qdrant-memory/scripts/
+cp scripts/backfill_memory.py /root/.openclaw/workspace/skills/qdrant-memory/scripts/
+
+# Restart the watcher
+python3 /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py --daemon
+
+# Verify
+ps aux | grep realtime_qdrant_watcher
+```
+
+### Backfilling Historical Memories (Optional)
+
+```bash
+python3 /root/.openclaw/workspace/skills/qdrant-memory/scripts/backfill_memory.py
+```
+
+### Verifying Your Upgrade
+
+```bash
+# 1. Check watcher is running
+ps aux | grep realtime_qdrant_watcher
+
+# 2. Verify source is "true-recall-base"
+curl -s "http://10.0.0.40:6333/collections/memories_tr/points/scroll"   -H "Content-Type: application/json" -d '{"limit": 1}' | jq '.result.points[0].payload.source'
+
+# 3. Check date coverage
+curl -s "http://10.0.0.40:6333/collections/memories_tr/points/scroll"   -H "Content-Type: application/json" -d '{"limit": 10000}' | jq '[.result.points[].payload.date] | unique | sort'
+```
+
+Expected output:
+- Source: `"true-recall-base"`
+- Dates: Array from oldest to newest memory

install.sh

@@ -96,3 +96,32 @@ echo "  curl -s http://$QDRANT_IP/collections/memories_tr | jq '.result.points_c
 echo ""
 echo "View logs:"
 echo "  sudo journalctl -u mem-qdrant-watcher -f"
+
+
+echo ""
+echo "=========================================="
+echo "UPGRADING FROM OLDER VERSION"
+echo "=========================================="
+echo ""
+echo "If you already have TrueRecall Base installed:"
+echo ""
+echo "1. Stop the watcher:"
+echo "   pkill -f realtime_qdrant_watcher"
+echo ""
+echo "2. Backup current files:"
+echo "   cp /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py \"
+echo "      /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py.bak"
+echo ""
+echo "3. Copy updated files:"
+echo "   cp watcher/realtime_qdrant_watcher.py \"
+echo "      /root/.openclaw/workspace/skills/qdrant-memory/scripts/"
+echo "   cp scripts/backfill_memory.py \"
+echo "      /root/.openclaw/workspace/skills/qdrant-memory/scripts/"
+echo ""
+echo "4. Restart watcher:"
+echo "   python3 /root/.openclaw/workspace/skills/qdrant-memory/scripts/realtime_qdrant_watcher.py --daemon"
+echo ""
+echo "5. Verify:"
+echo "   ps aux | grep realtime_qdrant_watcher"
+echo ""
+echo "For full upgrade instructions, see README.md"

scripts/backfill_memory.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+"""Backfill memory files to Qdrant memories_tr collection."""
+
+import os
+import json
+from datetime import datetime
+
+QDRANT_URL = "http://10.0.0.40:6333"
+MEMORY_DIR = "/root/.openclaw/workspace/memory"
+
+def get_memory_files():
+    """Get all memory files sorted by date."""
+    files = []
+    for f in os.listdir(MEMORY_DIR):
+        if f.startswith("2026-") and f.endswith(".md"):
+            date = f.replace(".md", "")
+            files.append((date, f))
+    return sorted(files, key=lambda x: x[0])
+
+def backfill_file(date, filename):
+    """Backfill a single memory file to Qdrant."""
+    filepath = os.path.join(MEMORY_DIR, filename)
+    with open(filepath, 'r') as f:
+        content = f.read()
+    
+    # Truncate if too long for payload
+    payload = {
+        "content": content[:50000],  # Limit size
+        "date": date,
+        "source": "memory_file",
+        "curated": False,
+        "role": "system",
+        "user_id": "rob"
+    }
+    
+    # Add to Qdrant
+    import requests
+    point_id = hash(f"memory_{date}") % 10000000000
+    resp = requests.post(
+        f"{QDRANT_URL}/collections/memories_tr/points",
+        json={
+            "points": [{
+                "id": point_id,
+                "payload": payload
+            }],
+            "ids": [point_id]
+        }
+    )
+    return resp.status_code == 200
+
+def main():
+    files = get_memory_files()
+    print(f"Found {len(files)} memory files to backfill")
+    
+    count = 0
+    for date, filename in files:
+        print(f"Backfilling {filename}...", end=" ")
+        if backfill_file(date, filename):
+            print("✓")
+            count += 1
+        else:
+            print("✗")
+    
+    print(f"\nBackfilled {count}/{len(files)} files")
+
+if __name__ == "__main__":
+    main()

session.md

@@ -19,7 +19,7 @@ true-recall-base (REQUIRED FOUNDATION)
     │   ├── Curator extracts atomic gems
     │   └── Plugin injects gems as context
     │
-    └──▶ true-recall-blocks (OPTIONAL ADDON)
+    └──▶ openclaw-true-recall-blocks (OPTIONAL ADDON)
         ├── Topic clustering
         └── Block-based retrieval
 ```
@@ -82,4 +82,4 @@ curl -s http://10.0.0.40:6333/collections/memories_tr | jq '.result.points_count
 
 ---
 
-*Next: Install true-recall-gems OR true-recall-blocks (not both)*
+*Next: Install true-recall-gems OR openclaw-true-recall-blocks (not both)*

watcher/realtime_qdrant_watcher.py

@@ -1,11 +1,14 @@
 #!/usr/bin/env python3
 """
-TrueRecall v1.2 - Real-time Qdrant Watcher
+TrueRecall v1.3 - Real-time Qdrant Watcher
 Monitors OpenClaw sessions and stores to memories_tr instantly.
 
 This is the CAPTURE component. For curation and injection, install v2.
 
 Changelog:
+- v1.3: Fixed crash loop (2551 restarts/24h) from FileNotFoundError on deleted session files.
+        Added chunking for long content (6000 char chunks) to prevent embedding token overflow.
+        Improved error handling for session file lifecycle.
 - v1.2: Fixed session rotation bug - added inactivity detection (30s threshold)
         and improved file scoring to properly detect new sessions on /new or /reset
 - v1.1: Added 1-second mtime polling for session rotation
@@ -27,7 +30,7 @@ from typing import Dict, Any, Optional, List
 # Config
 QDRANT_URL = os.getenv("QDRANT_URL", "http://10.0.0.40:6333")
 QDRANT_COLLECTION = os.getenv("QDRANT_COLLECTION", "memories_tr")
-OLLAMA_URL = os.getenv("OLLAMA_URL", "http://10.0.0.10:11434")
+OLLAMA_URL = os.getenv("OLLAMA_URL", "http://localhost:11434")
 EMBEDDING_MODEL = os.getenv("EMBEDDING_MODEL", "snowflake-arctic-embed2")
 USER_ID = os.getenv("USER_ID", "rob")
 
@@ -94,49 +97,124 @@ def clean_content(text: str) -> str:
     return text.strip()
 
 
+def chunk_text(text: str, max_chars: int = 6000, overlap: int = 200) -> list:
+    """Split text into overlapping chunks for embedding.
+    
+    Args:
+        text: Text to chunk
+        max_chars: Max chars per chunk (6000 = safe for 4K token limit)
+        overlap: Chars to overlap between chunks
+    
+    Returns:
+        List of chunk dicts with 'text' and 'chunk_index'
+    """
+    if len(text) <= max_chars:
+        return [{'text': text, 'chunk_index': 0, 'total_chunks': 1}]
+    
+    chunks = []
+    start = 0
+    chunk_num = 0
+    
+    while start < len(text):
+        end = start + max_chars
+        
+        # Try to break at sentence boundary
+        if end < len(text):
+            # Look for paragraph break first
+            para_break = text.rfind('\n\n', start, end)
+            if para_break > start + 500:
+                end = para_break
+            else:
+                # Look for sentence break
+                for delim in ['. ', '? ', '! ', '\n']:
+                    sent_break = text.rfind(delim, start, end)
+                    if sent_break > start + 500:
+                        end = sent_break + 1
+                        break
+        
+        chunk_text = text[start:end].strip()
+        if len(chunk_text) > 100:  # Skip tiny chunks
+            chunks.append(chunk_text)
+            chunk_num += 1
+        
+        start = end - overlap if end < len(text) else len(text)
+    
+    # Add metadata to each chunk
+    total = len(chunks)
+    return [{'text': c, 'chunk_index': i, 'total_chunks': total} for i, c in enumerate(chunks)]
+
+
 def store_to_qdrant(turn: Dict[str, Any], dry_run: bool = False) -> bool:
+    """Store a conversation turn to Qdrant, chunking if needed.
+    
+    For long content, splits into multiple chunks (no data loss).
+    Each chunk gets its own point with chunk_index metadata.
+    """
     if dry_run:
         print(f"[DRY RUN] Would store turn {turn['turn']} ({turn['role']}): {turn['content'][:60]}...")
         return True
     
-    vector = get_embedding(turn['content'])
+    content = turn['content']
-    if vector is None:
+    chunks = chunk_text(content)
-        print(f"Failed to get embedding for turn {turn['turn']}", file=sys.stderr)
-        return False
     
-    payload = {
+    if len(chunks) > 1:
-        "user_id": turn.get('user_id', USER_ID),
+        print(f"  📦 Chunking turn {turn['turn']}: {len(content)} chars → {len(chunks)} chunks", file=sys.stderr)
-        "role": turn['role'],
-        "content": turn['content'],
-        "turn": turn['turn'],
-        "timestamp": turn.get('timestamp', datetime.now(timezone.utc).isoformat()),
-        "date": datetime.now(timezone.utc).strftime('%Y-%m-%d'),
-        "source": "openclaw-true-recall-base",
-        "curated": False
-    }
     
-    # Generate deterministic ID
     turn_id = turn.get('turn', 0)
-    hash_bytes = hashlib.sha256(f"{USER_ID}:turn:{turn_id}:{datetime.now().strftime('%H%M%S')}".encode()).digest()[:8]
+    base_time = datetime.now().strftime('%H%M%S')
-    point_id = int.from_bytes(hash_bytes, byteorder='big') % (2**63)
+    all_success = True
     
-    try:
+    for chunk_info in chunks:
-        response = requests.put(
+        chunk_text_content = chunk_info['text']
-            f"{QDRANT_URL}/collections/{QDRANT_COLLECTION}/points",
+        chunk_index = chunk_info['chunk_index']
-            json={
+        total_chunks = chunk_info['total_chunks']
-                "points": [{
+        
-                    "id": abs(point_id),
+        # Get embedding for this chunk
-                    "vector": vector,
+        vector = get_embedding(chunk_text_content)
-                    "payload": payload
+        if vector is None:
-                }]
+            print(f"Failed to get embedding for turn {turn['turn']} chunk {chunk_index}", file=sys.stderr)
-            },
+            all_success = False
-            timeout=30
+            continue
-        )
+        
-        response.raise_for_status()
+        # Payload includes full content reference, chunk metadata
-        return True
+        payload = {
-    except Exception as e:
+            "user_id": turn.get('user_id', USER_ID),
-        print(f"Error writing to Qdrant: {e}", file=sys.stderr)
+            "role": turn['role'],
-        return False
+            "content": chunk_text_content,  # Store chunk content (searchable)
+            "full_content_length": len(content),  # Original length
+            "turn": turn['turn'],
+            "timestamp": turn.get('timestamp', datetime.now(timezone.utc).isoformat()),
+            "date": datetime.now(timezone.utc).strftime('%Y-%m-%d'),
+            "source": "true-recall-base",
+            "curated": False,
+            "chunk_index": chunk_index,
+            "total_chunks": total_chunks
+        }
+        
+        # Generate unique ID for each chunk
+        hash_bytes = hashlib.sha256(
+            f"{USER_ID}:turn:{turn_id}:chunk{chunk_index}:{base_time}".encode()
+        ).digest()[:8]
+        point_id = int.from_bytes(hash_bytes, byteorder='big') % (2**63)
+        
+        try:
+            response = requests.put(
+                f"{QDRANT_URL}/collections/{QDRANT_COLLECTION}/points",
+                json={
+                    "points": [{
+                        "id": abs(point_id),
+                        "vector": vector,
+                        "payload": payload
+                    }]
+                },
+                timeout=30
+            )
+            response.raise_for_status()
+        except Exception as e:
+            print(f"Error writing chunk {chunk_index} to Qdrant: {e}", file=sys.stderr)
+            all_success = False
+    
+    return all_success
 
 
 def is_lock_valid(lock_path: Path, max_age_seconds: int = 1800) -> bool:
@@ -332,10 +410,24 @@ def watch_session(session_file: Path, dry_run: bool = False):
     
     INACTIVITY_THRESHOLD = 30  # seconds - if no data for 30s, check for new session
     
-    with open(session_file, 'r') as f:
+    # Check file exists before opening (handles deleted sessions)
+    if not session_file.exists():
+        print(f"Session file gone: {session_file.name}, looking for new session...", file=sys.stderr)
+        return None
+    
+    # Track file handle for re-opening
+    try:
+        f = open(session_file, 'r')
+        f.seek(last_position)
+    except FileNotFoundError:
+        print(f"Session file removed during open: {session_file.name}", file=sys.stderr)
+        return None
+    
+    try:
         while running:
             if not session_file.exists():
                 print("Session file removed, looking for new session...")
+                f.close()
                 return None
             
             current_time = time.time()
@@ -346,6 +438,7 @@ def watch_session(session_file: Path, dry_run: bool = False):
                 newest_session = get_current_session_file()
                 if newest_session and newest_session != session_file:
                     print(f"Newer session detected: {newest_session.name}")
+                    f.close()
                     return newest_session
             
             # Check if current file is stale (no new data for threshold)
@@ -357,6 +450,7 @@ def watch_session(session_file: Path, dry_run: bool = False):
                         newest_session = get_current_session_file()
                         if newest_session and newest_session != session_file:
                             print(f"Current session inactive, switching to: {newest_session.name}")
+                            f.close()
                             return newest_session
                     else:
                         # File grew, update tracking
@@ -365,19 +459,31 @@ def watch_session(session_file: Path, dry_run: bool = False):
                 except Exception:
                     pass
             
-            # Process new lines and update activity tracking
+            # Check if file has grown since last read
-            old_position = last_position
+            try:
-            process_new_lines(f, session_name, dry_run)
+                current_size = session_file.stat().st_size
+            except Exception:
+                current_size = 0
             
-            # If we processed new data, update activity timestamp
+            # Only process if file has grown
-            if last_position > old_position:
+            if current_size > last_position:
-                last_data_time = current_time
+                old_position = last_position
-                try:
+                process_new_lines(f, session_name, dry_run)
-                    last_file_size = session_file.stat().st_size
+                
-                except Exception:
+                # If we processed new data, update activity timestamp
-                    pass
+                if last_position > old_position:
+                    last_data_time = current_time
+                    last_file_size = current_size
+            else:
+                # Re-open file handle to detect new writes
+                f.close()
+                time.sleep(0.05)  # Brief pause before re-opening
+                f = open(session_file, 'r')
+                f.seek(last_position)
             
             time.sleep(0.1)
+    finally:
+        f.close()
     
     return session_file