review-memory

A command for reviewing proposed changes to stored memories before they are applied.

In plain words
What is it for?
Use it to inspect maintenance proposals by memory area, compare the evidence and suggested changes, and approve, reject, edit, or promote items.
Why use it?
It keeps decisions such as merging duplicates, creating summaries, or removing memories under explicit human control. Unreviewed proposals remain pending rather than being approved automatically.

Command

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add commands/wuesteon/lean-memory/review-memory
Clone the repo
git clone --depth 1 https://github.com/Wuesteon/lean-memory
Per session 19 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 535 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00019 $0.00535
Opus 5 $0.00010 $0.00267
Sonnet 5 $0.00004 $0.00107
Haiku 4.5 $0.00002 $0.00053

Measured 2d ago against content hash 82189f81d0c8, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

review-memory scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugin/commands/review-memory.md · 46 lines

What it actually says

Maintenance runs offline and STAGES judgment calls (near-duplicate merges, summaries, evictions) as proposals; nothing changes in memory until a human approves. Walk the user through the queue and record only their explicit verdicts.

Hard rule — you may NOT decide any proposal on your own. Every approve / reject / edit / promote MUST come from an explicit user verdict in this conversation. If the user has not ruled on an item, leave it pending. A batch verb ("approve all exact dedups", "reject every eviction") is valid ONLY when the user says it in those words — never infer a batch decision from a single example or from silence. Silence is not consent; unreviewed proposals expire on their own.

Ask the user which namespace to review if they have not said. Then:

  1. Fetch the queue. Call the memory_review_queue MCP tool (namespace=<ns>, optional kind in dedup_near|summarize|evict, limit). It returns proposals grouped by subject entity, each carrying its evidence payload.

  2. Present, batched by entity and kind, with evidence. Show the before/after texts, cosine for near-dups, source facts for summaries, and the value signals for evictions. Keep it scannable so the user can rule quickly.

  3. Collect explicit verdicts. Ask when a verdict is missing or ambiguous. Do not proceed on an item the user has not decided.

  4. Decide per item the user ruled on. Call memory_review_decide (namespace, proposal_id, decision, optional edited_text) with decision in approve|reject|edit|promote. edited_text applies only to a summarize proposal the user chose to reword before approving. promote is for lifting an eviction's fact back to the hot tier.

  5. Summarize. Report what was applied, what was rejected, and what remains pending (untouched, still awaiting the user).

Notes for the user:

  • Approving is the only thing that changes stored memory. Rejecting leaves the spine byte-identical; ignoring a proposal lets it expire on its own.
  • Use one namespace per project/session — the same guidance as the other memory commands.
Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 46 lines · 19 tokens per session scan A 82189f81d0c8

Subscribe to this mod's changes

review-memory is a command published in the GitHub repository Wuesteon/lean-memory (47 stars, last pushed 26d ago), licensed Apache-2.0. It adds 19 tokens to every session and 535 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.