Self-improvement consolidation: slices 2–5¶
This page documents the second product batch in the OpenClaw Mem self-improvement consolidation line. The batch expands the read-only/staged-only skeleton after the initial surface validator and goal status pilot.
What this batch adds¶
Slice 2 — Goal survival pack seam¶
openclaw-mem goal pack builds a ContextPack-compatible fragment from a goal/controller receipt.
openclaw-mem goal pack --file goal.json --json
This is the goal-primitive alias for the existing active-line context pack shape. It is read-only and returns writes_performed=false.
Use it when an unfinished goal needs to survive compaction or handoff as a bounded pack fragment containing:
- goal id
- status
- objective
- current/next gate
- stop-loss state, when present
- verifier presence
Slice 3 — skill_capture staged proposal¶
openclaw-mem skill-capture propose lets an agent mark a learning candidate during a turn without editing live skills.
openclaw-mem skill-capture propose \
--text "When version changes, refresh uv.lock before CI." \
--target-skill ck-software-engineering-ops \
--rationale "Prevents lockfile freshness failures" \
--out .state/skill-capture/proposals/version-lock.json \
--json
Rules:
- writes only an explicit L1 staged proposal artifact
- never patches live skills mid-turn
- caps text size
- rejects path-like target skill names
- leaves curator judgment for a later review surface
Slice 4 — Skill Curator report-only review¶
openclaw-mem skill-curator review is a product-facing report-only alias for the existing self-curator skill review lane.
openclaw-mem skill-curator review --skill-root skills --json
It emits review-only lifecycle candidates and remains blocked until human review. By default it writes review JSON/Markdown artifacts under --out-root (default .state/skill-curator/runs/<timestamp>/); use --no-write for payload-only review. It does not apply patches.
Slice 5 — OpenClaw Mem System status¶
openclaw-mem mem-system status reports the current Store / Pack / Observe / Review / Curate surfaces in one read-only inventory.
openclaw-mem mem-system status --json
The status output includes:
- planes: Store, Pack, Observe, Review, Curate
- surface count
- state counts such as
stable,lab, andshadow writes_performed=falsetopology_changed=false
Non-goals¶
This batch does not enable:
- auto-continuation
- automatic skill mutation
- governed apply
- cron changes
- gateway/plugin configuration changes
- L3/L4 live mutation
Recommended next step¶
After these surfaces are merged and released, the next engineering line can prepare Slice 6: the staged mutation framework. That line should remain separate because it introduces proposal → apply → rollback semantics and touches higher-risk authority boundaries.