Openclaw Mem Curate

Curate Governance

Use scan → review → apply → verify → rollback as the explicit governed loop. Every wrapper preserves the complete engine receipt in inner.

Authority

• Scout with zero-write review before judgment.
• Keep low-risk unattended mutation within the explicit whitelist and mutation cap.
• Report post-hoc apply with exact changed paths and rollback receipt.
• Do not merge scout, governor, and writer authority into one hidden step.
• Run scheduled curation separately from heartbeat health checks.

Operate by verb

openclaw-mem curate scan --target memory --json
openclaw-mem curate scan --target episodes --json
openclaw-mem curate scan --target skills --skill-root <workspace>/skills --json
openclaw-mem curate scan --target facts --source-root <workspace> --json
openclaw-mem curate review --target memory --from-file <recommendation.json> --json
openclaw-mem curate apply --target memory --from-file <governor.json> --json
openclaw-mem curate verify --target memory --run-dir <receipt-dir> --json
openclaw-mem curate rollback --target memory --receipt <rollback.json> --json

Skill mutation remains checkpointed and whitelist-only: pass an explicit self-curator plan to curate apply --target skills --plan <plan.json>. Unsupported verb-target pairs return stable zero-write receipts.

See legacy commands when maintaining older automation. Aliases remain callable and emit deprecation guidance; do not remove them during migration.

Verify

openclaw-mem curate scan --target skills --skill-root <workspace>/skills --json
python -m pytest tests/test_self_curator.py -q
git diff --check