| name | memory-review-proposals |
| version | 0.9.5 |
| description | Review pending Memory Layer replacement proposals interactively inside the LLM CLI; explain why each memory update was proposed, gather codebase proof when needed, and approve or reject only after explicit confirmation |
Memory Proposal Review Skill
Use this skill when:
- the user asks to review proposed memories, pending memory updates, or the curation review queue
- the agent needs to explain why a candidate memory would replace an existing memory
- the user wants to gather more evidence before approving or rejecting a pending proposal
Do not use this skill for:
- normal post-task remembering; use
memory-remember
- creating new captures or running curation from scratch
- broad TUI usage questions that do not involve replacement proposals
Script
List pending proposals:
go run ./.agents/skills/memory-layer/scripts/main.go review-proposals list --project <project-slug>
Show one proposal:
go run ./.agents/skills/memory-layer/scripts/main.go review-proposals show --project <project-slug> --id <proposal-id>
Resolve one proposal only after explicit confirmation:
go run ./.agents/skills/memory-layer/scripts/main.go review-proposals approve --project <project-slug> --id <proposal-id>
go run ./.agents/skills/memory-layer/scripts/main.go review-proposals reject --project <project-slug> --id <proposal-id>
Workflow
- List pending proposals and pick the next proposal the user wants to inspect.
- Explain the proposal in plain language:
- target memory summary
- candidate memory summary and full text
- candidate type, score, policy, and matcher reasons
- what would happen if approved
- If evidence is unclear, offer investigation options in the CLI conversation:
- search for proof in the codebase with
rg
- inspect the target memory history with
memory history <target-memory-id> --json
- search project memory with
memory query --project <project> --question "<question>" --json
- skip and leave the proposal pending
- For codebase proof, derive focused searches from the candidate text, target/candidate summaries, file paths in sources when visible, and the matcher reasons. Report both confirming and contradicting evidence.
- Recommend approve, reject, or leave pending. Prefer leave pending when evidence is insufficient.
- Ask for explicit confirmation before running approve or reject. Do not bulk-resolve proposals unless the user explicitly asks and each decision is justified.
- After resolution, refresh the proposal list and continue with the next selected proposal if requested.
Decision Rules
- Approve when the candidate is more current, more specific, and supported by code/docs/history evidence.
- Reject when the candidate is duplicate noise, stale, too broad, contradicted by the repo, or would replace a better existing memory.
- Leave pending when the evidence is not clear enough to make a durable-memory decision.
- Never invent provenance. If proof cannot be found, say that directly.
Runtime Requirement
This focused skill uses the shared Go helper under .agents/skills/memory-layer/scripts/.
go must be available on PATH for these helper commands to run.
