菜单
Two-tier memory management: write, update, prune, and search memories with semantic deduplication. Maintains MEMORY.md index under 200 lines — maintains the existing index; for writing one new structured entry instead, use full version `wisdom-garden:memory-garden`.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | memory-curator |
| description | Two-tier memory management: write, update, prune, and search memories with semantic deduplication. Maintains MEMORY.md index under 200 lines — maintains the existing index; for writing one new structured entry instead, use full version `wisdom-garden:memory-garden`. |
| triggers | ["manage memories","curate memory","prune memories","memory health","organize memories"] |
| version | 1.0.1 |
| model | sonnet |
| category | remember-continue |
Manages the persistent memory system — writes, updates, prunes, and deduplicates memories while keeping the MEMORY.md index lean and navigable.
Memory degrades in two directions: it grows bloated with stale or duplicate entries until the index is unnavigable, or it gets pruned too aggressively and the agent forgets corrections that cost real sessions to learn. This skill walks the middle path — surgically removing what's derivable from code or superseded by later facts, while protecting user preferences, feedback corrections, and hard-won patterns. A healthy memory store is a compressed model of the project, not a transcript of every session.
Do NOT run aggressively after a single session without a convergence signal — premature pruning removes memories before they've proven their value.
Read MEMORY.md index and all referenced memory files in the memory directory. Build an inventory of:
Detect semantically duplicate memories — same fact stored with different wording. Criteria:
Identify memories referencing things that no longer exist:
Combine related memories that cover the same topic:
Remove memories that are:
Rebuild MEMORY.md:
Output summary:
Index is at 187 lines (GROWING). Running the curator:
src/legacy/payments.js which no longer exists in the repoconvergence-checker — triggers memory-curator runs based on session count and index size thresholdsseed-curator — manages the seed library, which lives alongside the memory system; coordinate when the boundary between seeds and memories is blurryProduces markdown memory artifacts (conversation summaries, seed files, philosophical reflections, doc updates) and a dated compression log by distilling a long conversation into its essential decisions and learnings. For the lighter, routine end-of-session wrap-up, use `session-compression` instead — this skill is for a long conversation that needs the fuller multi-artifact treatment. Use when: "compress this context", "distill this conversation", "create a memory artifact", "condense this history", "extract key wisdom before handoff".
Produces a dispatch plan — isolation model, agent count, sequencing, and model assignments — for multi-agent parallel work. Plans the mechanics once you already know you're doing a parallel dispatch. If unsure which orchestration pattern fits at all, use `orchestration-pattern-selector` first; if the shape is a single conductor coordinating specialists, use `maestro-orchestration` directly. Use when: "dispatch agents in parallel", "run multiple tracks simultaneously", "plan parallel agent strategy", "should this be foreground or background".
Conductor agent pattern that decomposes complex tasks, dispatches specialist sub-agents, manages dependencies, and synthesizes results into a unified deliverable. One specific orchestration pattern — a single conductor coordinating specialists. Use directly when that shape already fits; use `orchestration-pattern-selector` first if unsure, or `agent-dispatch-playbook` to plan a general parallel dispatch's mechanics. Use when: 'coordinate multiple specialist agents', 'orchestrate a complex multi-phase task', 'run a conductor pattern across agents', 'this task needs live agent coordination', 'decompose and dispatch to sub-agents'.
Selects the correct multi-agent orchestration pattern for a task using an 11-signal selection matrix. Use this first, upstream of `agent-dispatch-playbook` (which plans dispatch mechanics once the shape is known) and `maestro-orchestration` (one specific pattern this skill might select). Use when choosing which orchestration pattern fits a multi-agent task.
Produces a verified multi-track execution by dispatching parallel agents with file manifests, independent verification, and status tracking. Executes an already-planned split — the execution-side counterpart to `specification-driven-development:parallel-tracks`, which plans the phased structure first. Use when: "dispatch agents", "parallel tracks", "multi-track work", "fan out", "divide and conquer".
Produces a structured codebase map — directory roles, entry points, dependency graph, architectural patterns, and dragon locations — for any unfamiliar repository. Canonical version (adds a defined reading-order output); `dojo-craft:codebase-viewer` is the single-file lean variant with the same core output. Use when: 'map this codebase for me', 'I need to understand this repo before I start', 'give me a reading order for this project', 'where are the dragons in this code'.