CLI: reinforcement and validation
These commands operate the self-maintaining memory system: activation scores earned through real usage, and the validation pipeline that re-checks hot memories against project evidence.
| Command | Writes | Purpose |
|---|---|---|
memory scores | No | Show activation scores, volatility, and review flags for a project. |
memory validate | Yes | Run evidence-backed validation for one memory (LLM call). |
memory review | Yes | List, apply, or reject corrections proposed by validation. |
memory scores
Read-only view of how strongly each memory has been used recently, with time decay applied.
memory scores --project memory
memory scores --project memory --needs-review
memory scores --project memory --limit 50 --jsonColumns worth knowing: ACTIVATION is the decay-corrected score (crossing the configured threshold makes the memory due for validation), VOLAT is the observed change rate of the memory's source files, and memories flagged by validation show [NEEDS REVIEW] and rank lower in query results until resolved.
memory validate
Triggers the validation pipeline for one memory: gather evidence (sources, provenance, related memories, recent git history), get an LLM verdict, apply the configured policy. Manual runs bypass the activation threshold but respect the global daily budget.
memory validate <memory-id> # uses the configured dry-run default
memory validate <memory-id> --dry-run # report only, never changes anything
memory validate <memory-id> --execute # allow policy actions to applyOutcomes: revalidated (accurate and clear), reworded (wording improved as a new version — only with auto-apply enabled), correction_pending (proposal queued for review), or flagged_needs_review (weak or contradictory evidence; content never modified). Dry runs report the same outcomes as would_* actions.
Memories captured with --file-changed provenance validate best: the pipeline can check their source files and git history. Memories without file sources usually come back unsupported — that is the system refusing to guess, not a failure.
memory review
Corrections proposed by validation wait here for a human decision.
memory review list --project memory # pending corrections
memory review list --project memory --all # every validation run
memory review apply <run-id> # write the proposal as a new memory version
memory review reject <run-id> # record the rejectionBoth resolutions clear a standing needs-review flag. Applying creates a new immutable version under the same canonical memory, so memory history shows the full chain and any change can be reverted.
Not to be confused with memory proposals, which reviews curation replacement proposals (new captures that overlap existing memories). memory review handles validation corrections (existing memories found outdated or unclear).