Persistent memory for OpenClaw. Fully local, zero vendor dependency.
- Verbatim Search — SQLite FTS5 over session transcripts. Exact-match search across your entire conversation history. This is the primary, actively maintained feature.
- Semantic Memory (optional) — LanceDB + Ollama plugin. Auto-captures facts, preferences, and decisions, and auto-recalls relevant context before each response.
┌─────────────────────────────────────────────────────┐
│ Memento │
│ │
│ ┌───────────────────────┐ ┌──────────────────┐ │
│ │ Verbatim Search │ │ Semantic Memory │ │
│ │ (primary) │ │ (optional) │ │
│ │ │ │ │ │
│ │ SQLite + FTS5 │ │ LanceDB │ │
│ │ + JSONL indexer │ │ + Ollama │ │
│ │ │ │ (nomic-embed) │ │
│ │ "Find every message │ │ "What was that │ │
│ │ containing the word │ │ thing about │ │
│ │ 'migration'" │ │ budgets?" │ │
│ └───────────────────────┘ └──────────────────┘ │
│ │
│ Verbatim ≈ exact search Semantic ≈ fuzzy recall │
└─────────────────────────────────────────────────────┘
- OpenClaw installed and running
- Python 3.9+
- macOS or Linux
# Clone this repo (or just grab scripts/transcript-search.py)
git clone https://github.com/jdawe/memento.git
cd memento
# Run initial index
python3 scripts/transcript-search.py index
python3 scripts/transcript-search.py statsThe script reads environment variables (with sensible defaults):
| Variable | Default | Description |
|---|---|---|
MEMENTO_WORKSPACE_DIR |
~/.openclaw/workspace |
Your OpenClaw workspace path |
MEMENTO_SESSIONS_DIR |
~/.openclaw/agents/main/sessions |
Path to OpenClaw JSONL session files |
MEMENTO_DB_PATH |
$MEMENTO_WORKSPACE_DIR/data/transcripts.db |
Path to the SQLite database |
Memento indexes OpenClaw agent sessions only — it operates at the harness level, not the
underlying model CLI. If you drive a model directly (e.g. claude in a terminal, outside
OpenClaw), that transcript isn't indexed.
Override them if your setup differs:
MEMENTO_SESSIONS_DIR=/custom/path MEMENTO_DB_PATH=/custom/db.sqlite python3 scripts/transcript-search.py index# Search all conversations
python3 scripts/transcript-search.py search "budget spreadsheet"
# Filter by role and date range
python3 scripts/transcript-search.py search "API migration" --role assistant --after 2026-03-01
# Full message content (no truncation)
python3 scripts/transcript-search.py search "architecture" --full --limit 5
# Rebuild from scratch
python3 scripts/transcript-search.py reindex
# Remove cross-session duplicate messages (content-hash based)
python3 scripts/transcript-search.py dedup
# Show stats
python3 scripts/transcript-search.py statsKeep the index fresh with a cron job:
# Via OpenClaw cron (recommended):
openclaw cron add transcript-indexer --every 5m \
--message "Run: bash /path/to/memento/scripts/transcript-index-cron.sh"
# Or via system crontab:
*/5 * * * * /path/to/memento/scripts/transcript-index-cron.sh| Component | What it does |
|---|---|
| JSONL parser | Reads OpenClaw session transcripts (user, assistant, tool messages) |
| SQLite FTS5 | Full-text search index with BM25 ranking |
| Incremental indexing | Only processes new/changed files (tracks by file size) |
| Cross-session dedup | Content-hash based; run dedup to collapse duplicate messages across sources |
| Triggers | FTS index auto-syncs on insert/update/delete |
Disabled by default. Enable it if you want automatic fact/preference capture and recall on top of verbatim search.
brew install ollama # macOS
# or: curl -fsSL https://ollama.com/install.sh | sh # Linux
ollama serve & # starts local API on 127.0.0.1:11434
ollama pull nomic-embed-text # 274MB, 768-dim, fast on Apple SiliconCopy examples/openclaw-plugin-config.json into your ~/.openclaw/openclaw.json (merge with existing config):
{
"plugins": {
"memory": "memory-lancedb",
"entries": {
"memory-lancedb": {
"enabled": true,
"config": {
"embedding": {
"apiKey": "ollama",
"model": "nomic-embed-text",
"baseURL": "http://127.0.0.1:11434/v1"
},
"autoCapture": true,
"autoRecall": true
}
}
}
}
}
⚠️ Use127.0.0.1, notlocalhost. Node.js may resolvelocalhostto IPv6 (::1), but Ollama only listens on IPv4. This causes silent connection failures.
openclaw gateway restart
openclaw status
# Look for: Memory │ enabled (plugin memory-lancedb)| Feature | Description |
|---|---|
| Auto-capture | Extracts facts, preferences, decisions from every conversation and embeds them |
| Auto-recall | Searches stored memories before each response, injects relevant context |
| Manual tools | memory_store — save explicitly. memory_recall — search explicitly |
Data lives at ~/.openclaw/memory/lancedb/ (~1-5 MB per thousand memories).
| Question | Layer | Why |
|---|---|---|
| "What did we decide about the API?" | Semantic (LanceDB) | Fuzzy match, context-aware |
| "Find the exact message where I said 'migration'" | Verbatim (SQLite) | Precise, grep-like |
| "Do I prefer dark mode?" | Semantic (LanceDB) | Preference recall |
| "Show me everything from last Tuesday" | Verbatim (SQLite) | Date-filtered exact search |
Verbatim search is on-demand — run it when you need exact quotes or historical context. Semantic memory (if enabled) is automatic — it captures and recalls without you doing anything.
| Component | Size |
|---|---|
| Ollama binary | ~32 MB |
| nomic-embed-text model | ~274 MB |
| LanceDB data | ~1-5 MB per 1K memories |
| SQLite transcript DB | ~5-15 MB per 10K messages |
Fully local. No API calls. No cloud dependency.
MIT