Skip to content

Evaluator path

Use this page when you want the shortest honest route through openclaw-mem.

If you have 5 minutes

Goal: prove the core mechanism on synthetic memory only.

git clone https://github.com/phenomenoner/openclaw-mem.git
cd openclaw-mem
uv sync --locked
uv run --python 3.13 --frozen -- \
  python benchmarks/trust_policy_synthetic_proof.py --json

Expected result:

  • passed: true
  • one quarantined row selected by vanilla packing is excluded by trust-aware packing
  • selected rows keep citation coverage
  • the trust policy explains the exclusion reason

Read next: Trust-policy synthetic proof.

If you have 30 minutes

Goal: decide whether the sidecar-first adoption path fits your operator workflow.

  1. Run the trust-policy synthetic proof.
  2. Run the 60-second reality proof.
  3. Read Core vs Advanced Labs.
  4. Read Choose an install path.
  5. Skim Automation status so you know what is automatic, opt-in, partial, or roadmap.

Decision point:

  • If you only need a short chat memory, stop here.
  • If you need inspectable agent memory with citations and rollback posture, try the sidecar path.
  • If you need live-turn recall orchestration, evaluate the optional mem engine after sidecar proof.

If you have one afternoon

Goal: evaluate whether openclaw-mem belongs in a real OpenClaw operator stack.

  1. Run the 5-minute proof and save the JSON receipt.
  2. Install the sidecar using Quickstart and Install modes.
  3. Ingest a small sanitized test memory file, not production memory.
  4. Run search, timeline, get, and pack on 3–5 realistic questions.
  5. Inspect citations and trace receipts for at least one expected include and one expected exclude.
  6. Only then consider optional mem-engine promotion.

Keep Advanced Labs out of the first evaluation unless your use case explicitly needs them.