AGENT SKILL

memory-dream

Reach for this when a project's auto-memory has accumulated rot a consolidation pass would catch, or when you want to measure whether a memory change actually improved recall.

License 许可
MIT
最近更新
2026-11-02
Stars
★ 1
Forks
0

01 前言 Front Matter

name
memory-dream
description
Reach for this when a project's auto-memory has accumulated rot a consolidation pass would catch, or when you want to measure whether a memory change actually improved recall.

03 SKILL.md 正文

memory-dream: the sleep cycle for Claude Code auto-memory

Auto-memory grows monotonically. Sessions write notes and nothing ever consolidates them, so notes supersede each other and both survive, descriptions go stale and route future sessions to wrong answers, long logs bury the one durable lesson, and eventually MEMORY.md outgrows the index load cap and the newest notes stop routing at all. memory-dream is the consolidation pass that was missing — and the eval that proves the pass helped rather than assuming it.

This skill orients; it does not act. The work lives in two commands, each operator-gated by design. When the situation below appears, surface the right command to the operator and let them run it — consolidating memory by hand, or calling the apply step directly, bypasses every gate the pass exists for.

Reach for a command

  • /memory-dream:dream — a full consolidation pass over live memory: deterministic triage finds the rot, a restricted subagent drafts fixes it cannot apply, the operator approves a diff item by item, and only approved changes are written (every one recoverable). Suggest this when triage would find flagged notes — its checks are mechanical: supersession markers, oversized bodies, decayed confidence, stale-dated content, or an index over the load cap (which the pass's archive tier demotes rather than deletes).
  • /memory-dream:eval — a read-only recall-regression eval: it freezes a content-anchored question suite over the live corpus and scores whether a fixed judge, seeing only MEMORY.md descriptions, routes each question to a note that answers it. Suggest this to baseline recall before a pass and to measure the after delta, or any time the question is "did a memory change actually improve recall?"

Both commands are disable-model-invocation: true: the operator invokes them, this skill only points the way.

The invariant that makes it safe

Every semantic change is proposal-only. The pass drafts, verifies, previews, and applies solely what the operator approved by token, item by item — it treats note bodies as untrusted data throughout, and the drafting subagent has no mutating tools, so a prompt injection inside a note cannot drive a write. Both agents declare an explicit minimal tool list (Glob only, unused by their prompts); an empty list would grant every inherited tool. All task content is supplied inline, and omitClaudeMd: true excludes unrelated host instructions. Apply snapshots every touched file first, and memory-dream restore reverses a pass. Read SECURITY.md for the trust model before installing on a shared machine.

Beyond the plugin

memory-dream is also a standalone, stdlib-only Python CLI that works from a clone alone — useful for a dry run before installing the plugin, or for driving the pipeline outside Claude Code:

python3 -m memory_dream doctor     # preflight: what works on this machine
python3 -m memory_dream triage     # what a pass would find here, read-only

Start with doctor; it reports the operating mode, whether the consent trace can see your transcripts, and which caps apply. Full documentation, the gate stack, and every tunable threshold are in the repository README and docs/.