OpenClaw Mem

Sidecar-first memory capture for OpenClaw.

This ClawHub bundle installs the lightweight openclaw-mem observation plugin, not the optional memory-slot engine. It is the safest default adoption path: capture OpenClaw activity into local JSONL, ingest it into the openclaw-mem SQLite sidecar, and keep your current OpenClaw memory backend untouched.

What this bundle includes

• extensions/openclaw-mem/, the sidecar capture plugin.
• Operator docs for setup, deployment, auto-capture, and the optional engine path.
• No forced memory-slot migration.
• No schema migration.
• No remote service requirement for the sidecar path.

What the sidecar does

• Captures tool results into append-only JSONL for later ingest.
• Optionally spools bounded episodic events, including conversation/tool/ops events.
• Redacts common secrets by default before persistence.
• Annotates memory-tool activity so retrieval and backend behavior remain auditable.
• Lets you run local recall, timeline, graph, drift, provenance, and optimization review workflows from the GitHub repo.

Install from ClawHub

openclaw bundles install clawhub:openclaw-mem-lyria

Then enable the plugin entry in your OpenClaw config and restart/reload your gateway using your normal OpenClaw deployment path.

Minimal config shape:

{
  "plugins": {
    "entries": {
      "openclaw-mem": {
        "enabled": true,
        "config": {
          "outputPath": "memory/openclaw-mem-observations.jsonl",
          "captureMessage": false,
          "redactSensitive": true,
          "backendMode": "auto",
          "annotateMemoryTools": true,
          "excludeAgents": ["cron-watchdog", "healthcheck"],
          "episodes": {
            "enabled": false,
            "outputPath": "memory/openclaw-mem-episodes.jsonl"
          }
        }
      }
    }
  }
}

Recommended adoption path

1. Install this sidecar bundle. Keep your existing OpenClaw memory slot.
2. Capture to local JSONL. Start with captureMessage: false and redactSensitive: true.
3. Run a local proof from GitHub. Clone the repo and ingest a small sample into SQLite.
4. Schedule harvest / review only after the proof passes. Keep outputs in directories you control.
5. Consider the engine later. Only promote to openclaw-mem-engine when you want it to own the OpenClaw memory slot.

Local proof, no OpenClaw config change

git clone https://github.com/phenomenoner/openclaw-mem.git
cd openclaw-mem
uv sync --locked

DB=/tmp/openclaw-mem.sqlite
python3 ./scripts/make_sample_jsonl.py --out /tmp/openclaw-mem-sample.jsonl

uv run --python 3.13 --frozen -- python -m openclaw_mem --db "$DB" --json status
uv run --python 3.13 --frozen -- python -m openclaw_mem --db "$DB" --json ingest --file /tmp/openclaw-mem-sample.jsonl
uv run --python 3.13 --frozen -- python -m openclaw_mem --db "$DB" --json search "OpenClaw" --limit 5
uv run --python 3.13 --frozen -- python -m openclaw_mem --db "$DB" --json timeline 2 --window 2
uv run --python 3.13 --frozen -- python -m openclaw_mem --db "$DB" --json optimize review --limit 200

Engine path

The optional openclaw-mem-engine is intentionally not the default marketplace install path. It is a stronger posture because it can become the active memory-slot backend and enable hybrid recall / controlled automation.

Install it from GitHub after reading the docs:

• Repo: https://github.com/phenomenoner/openclaw-mem
• Engine docs: docs/mem-engine.md
• Deployment docs: docs/deployment.md
• Release notes: https://github.com/phenomenoner/openclaw-mem/releases

Rollback

Disable or remove the plugins.entries.openclaw-mem entry. Because this bundle is sidecar-only, your native memory backend remains untouched.

Version

This ClawHub release tracks openclaw-mem v1.9.1, including the runtime latency improvements in sidecar-adjacent retrieval workflows and the current sidecar plugin contract.