菜单
Use this skill when changing RECALL capture mode, when cleaning noisy memory, when repairing local storage, or when updating lifecycle state after review. Use proactively for explicit maintenance requests, never for ordinary retrieval or new-memory capture.
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
Use this skill proactively when creating a reusable category, when configuring retrieval weight, when refining an auto-created definition, or when organizing repeated memory types. Trigger when category configuration is needed; invoke automatically for durable taxonomy work, not one-off tags.
Use this skill when a fresh RECALL session or new thread needs the usage contract, when the user asks what RECALL remembers, when prior project decisions or commands are relevant, or when a durable requirement, risk, decision, command, or project-state update needs to be saved. Auto-loads at sessionStart on Kimi Code; invoke before other RECALL skills so provider provenance and local-only storage rules are established.
Use this skill when the user explicitly asks to run, judge, or compare a RECALL benchmark (e.g. "run a judged normal bench", "recall-bench light", "benchmark RECALL against the baseline"). It runs the deterministic bench harness, optionally acts as the LLM judge for emitted judge tasks, aggregates scores, and reports results. Never invoke for ordinary RECALL memory usage, plugin development, or test-suite runs.
Use this skill proactively when debugging a known failure, when implementing work shaped by prior decisions, when testing with verified commands, or when starting a new thread that needs project state. Trigger when local memory prevents repeated investigation; invoke automatically only when project context is relevant.
Use this skill when deciding whether candidate context belongs in Recall memory, repo docs, skill/plugin instructions, provider config, or current chat only; when planning memory cleanup; when detecting stale, duplicate, conflicting, low-value, source-backed, command, or weak-preference memories; or when applying safe non-destructive hygiene. Use proactively for memory quality policy before using manage-memory for direct ID-based mutation.
Use this skill proactively when debugging a durable failure, when testing a verified command, when implementing an accepted decision, or when recording a standing requirement for future threads. Trigger when durable evidence should persist; invoke automatically only for explicit evidence, never drafts.
| name | manage-memory |
| description | Use this skill when changing RECALL capture mode, when cleaning noisy memory, when repairing local storage, or when updating lifecycle state after review. Use proactively for explicit maintenance requests, never for ordinary retrieval or new-memory capture. |
Use this skill for RECALL maintenance and lifecycle control after memory already exists. RECALL is local-only project memory. Work only in the active project's RECALL memory directory: .recall/ for new projects, or existing .codex_memory/ stores for legacy projects. Do not store or repeat secrets.
Use this skill when the user asks to:
configure-capturereview-memory or audit-memory passDo not use this skill to create a brand-new durable insight from scratch. For new facts, use save-insight. For inspection-only requests, start with review-memory or retrieve-memory.
This skill usually receives one of these inputs:
| Input shape | Use it for | Example |
|---|---|---|
| broad cleanup request | capture mode or noise cleanup | "set recall to minimal mode" |
| memory id plus note | lifecycle updates | "mark memory 12 stale" |
| multiple ids | merge or supersede | "merge 18 19 20" |
| health complaint | repair or doctor | "memory index seems broken" |
If the user did not provide IDs and the state is unclear, retrieve the IDs first with review-memory or retrieve-memory.
Use recall_skill.py only:
Run these examples from the installed/source plugin root so ./scripts/recall_skill.py resolves. If the shell is in the active project repository, do not look for ./scripts there; use the adapter's absolute path and pass --root <project-root> when needed. Lifecycle, edit/delete, cleanup, and repair commands are intentionally grouped in this one skill instead of separate public skill folders.
python ./scripts/recall_skill.py configure-capture minimal
python ./scripts/recall_skill.py configure-capture off
python ./scripts/recall_skill.py archive-noise
python ./scripts/recall_skill.py doctor
python ./scripts/recall_skill.py repair
python ./scripts/recall_skill.py confirm-memory 12
python ./scripts/recall_skill.py resolve-memory 12 --note "Implemented."
python ./scripts/recall_skill.py stale-memory 12 --note "Needs reconfirmation."
python ./scripts/recall_skill.py supersede-memory 12 18 --note "Memory #18 replaces #12."
python ./scripts/recall_skill.py merge-memories 18 19 20 --note "Merged duplicates."
python ./scripts/recall_skill.py prune-memory 12 --note "Archive after review."
python ./scripts/recall_skill.py edit-memory 12 --summary "Corrected summary."
python ./scripts/recall_skill.py delete-memory 12 --confirm DELETE-12
Treat lower-level backend files as internal plumbing, not public workflow.
This skill receives an existing-memory maintenance request and returns the lifecycle, capture, or storage-health change that was made. It does not create new durable facts and it does not perform inspection-only reports except to collect IDs needed for a maintenance action.
Contract schema:
| Field | Required | Meaning |
|---|---|---|
intent | yes | capture-mode, cleanup, lifecycle, doctor, or repair |
ids | for lifecycle | exact memory IDs gathered from user input or review output |
note | for state change | concise reason preserved in lifecycle metadata |
safety | for deletion | explicit confirmation token for destructive delete |
result | yes | command run, changed IDs, status/capture mode, follow-up |
Use the contract asset when deciding whether the request belongs here:
{"asset":"assets/contract.json","kind":"maintenance-boundary"}
review-memory or retrieve-memory if IDs or current state are unclear.configure-capture manual|minimal|standard|off to control automatic hook writes.
The modes are enforced inside the hooks (not by agent goodwill):
standard: full automatic capture — per-tool evidence, prompt signals, stop notes, session summaries.minimal: no per-tool evidence buffering; prompt signals, stop notes, and session summaries still run.manual: only explicit cues (@recall remember this) and skill/MCP saves; no automatic hook capture.off: no hook capture of any kind, including explicit prompt cues (the hook explains how to re-enable);
skill and MCP saves stay available. Retrieval is governed separately by configure-recall.archive-noise for non-destructive cleanup of low-value automatic command history.doctor and repair only for local storage or index maintenance.
Use migrate-store (dry-run by default, --apply to execute) to move a
legacy .codex_memory/ store into provider-neutral .recall/: it copies
via the sqlite backup API, verifies record counts, rebuilds the index,
and leaves the legacy directory in place as a frozen backup.delete-memory only with explicit user intent. Prefer archival over deletion.Return a concise maintenance result:
When a command returns JSON, summarize the important fields instead of dumping raw output unless the user explicitly asked for it.
{
"action": "resolve-memory",
"id": 12,
"status": "resolved",
"follow_up": "none"
}
For multi-record work, keep the same shape and make changed IDs explicit:
{"action":"archive-noise","changed_ids":[18,19,22],"status":"archived","follow_up":"none"}
Broad cleanup after an audit:
python ./scripts/recall_skill.py archive-noise --apply --limit 50
Lifecycle follow-up after a fix shipped:
python ./scripts/recall_skill.py resolve-memory 12 --note "Implemented in release 0.1.1."
Storage health check when retrieval looks wrong:
python ./scripts/recall_skill.py doctor
python ./scripts/recall_skill.py repair
Capture-mode change:
python ./scripts/recall_skill.py configure-capture minimal
Safe lifecycle correction:
python ./scripts/recall_skill.py stale-memory 42 --note "Repository state now contradicts this record."
| User intent | Preferred command | Why |
|---|---|---|
| lower automatic write volume | configure-capture minimal | preserves useful automation with less noise |
| stop automatic writes entirely | configure-capture off | strongest suppression |
| clean noisy historical command spam | archive-noise | archives safely, does not delete |
| mark a memory outdated but keep it | stale-memory or prune-memory | preserves history |
| replace an older memory with a newer one | supersede-memory | records relationship |
| remove a wrong record completely | delete-memory | only with explicit user intent |
| Request | This skill action | Handoff |
|---|---|---|
| "save this decision" | decline direct write | skills/save-insight |
| "what does memory know?" | inspect only if IDs are needed | skills/review-memory |
| "find the old test command" | no lifecycle mutation | skills/retrieve-memory |
| "memory 12 is obsolete" | run stale/supersede/prune | none |
| "index looks broken" | run doctor, then repair if indicated | none |
review-memory first.doctor reports index problems, use repair before creating more memories.save-insight.archive-noise or prune-memory over delete-memory.See Review Memory for inspection and quality review before making changes. See Retrieve Memory for targeted lookup by query. See Save Insight for creating new durable memory instead of editing existing records. See Maintenance playbook for cleanup sequencing. Sibling routes: skills/review-memory, skills/retrieve-memory, skills/save-insight.