openclaw-mem¶
Trust-aware context packing for OpenClaw.
This docs site is organized around one core idea:
- keep prompt packs small
- keep selection inspectable
- keep trust tiers and provenance visible
- keep stale / untrusted / hostile content from quietly polluting future answers
This docs site is organized around a practical product journey:
- what problem is being solved,
- how trust-aware packing addresses it,
- where to verify the proof quickly,
- and how to adopt with a low-risk rollout path.
Two-plane framing:
- Query plane (default): recall + trust-aware context packing with citations and receipts.
- Action plane (optional): recommendation-only hygiene and maintenance review queues (no silent writeback to durable memory).
Recommended rollout sequence¶
- Prove it locally (5 minutes) — Trust-aware pack proof
- Run sidecar on existing OpenClaw (default) — Choose an install path
- Promote to optional mem engine when needed — Mem Engine reference
Start here¶
- Trust-aware pack proof — canonical before/after artifact
- About openclaw-mem — product scope, audience, and boundaries
- Quickstart — fastest local proof from a fresh clone
- Portable pack capsules —
openclaw-mem capsuleseal / inspect / verify / diff + canonicalexport-canonical+ boundedrestore(isolated/new target only) - Reality check & status — what is done, partial, and roadmap
Proof artifacts¶
Product shape¶
openclaw-mem is intentionally split into two layers:
- Sidecar (default) — capture, ingest, local recall, pack receipts, graph query/drift checks, hygiene review
- Mem Engine (optional) — memory-slot backend for hybrid recall and bounded automation
Start with the inspectable sidecar, then add the engine only when you need it.