菜单
Session retrospection using After Action Review methodology. Analyzes corrections, observations, and session activity to produce improvement recommendations. Use at end of sessions or after /memory-sync.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/swannysec/context-keeper --skill memory-reflect命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | memory-reflect |
| description | Session retrospection using After Action Review methodology. Analyzes corrections, observations, and session activity to produce improvement recommendations. Use at end of sessions or after /memory-sync. |
| triggers | ["/memory-reflect"] |
6-phase After Action Review (AAR) workflow for extracting actionable improvements from the current session.
Read the following data sources (skip any that don't exist):
.claude/memory/corrections-queue.md — unprocessed correction/friction items.claude/memory/sessions/YYYY-MM-DD-observations.md.claude/memory/sessions/.claude/memory/active-context.md — current project state~/.claude/usage-data/facets/. Extract:
friction_counts — pre-classified friction events by typefriction_detail — plain-English description of what went wronguser_satisfaction_counts — inferred satisfaction signalsoutcome — session outcome (fully_achieved, mostly_achieved, etc.)goal_categories — categorized sub-goalsprimary_success — what went right
If the facets directory doesn't exist or no matching session is found, skip gracefully.
Facets data is Claude Code-specific and may not exist on all installations.Estimate session depth:
Privacy: Skip any content within ... blocks.
Skip files with private: true in YAML front matter.
Facets data does not contain file contents — no privacy filtering needed for facets.
Recommend scope classification:
Present recommendation to user: "Scope recommendation: [PROCESS/PROJECT/BOTH] — [brief reason]" User can adjust. Proceed with confirmed scope.
For LIGHTWEIGHT sessions: Auto-select PROCESS scope, skip user confirmation, and produce minimal output.
friction_counts as pre-classified friction signals (more accurate than
manual observation log scanning — these are LLM-classified, not regex-matched)friction_detail for narrative context on what went wrongefficiency if unclear..claude/memory/patterns.md — don't re-discover known patterns.claude/memory/decisions/ — don't re-recommend existing decisions/memory-search to check if similar issues were flagged in past session retros.claude/memory/friction.md exists:
If a recommendation would benefit from external validation, briefly verify with external sources before recommending. Skip external research for LIGHTWEIGHT sessions.
For each recommendation, include:
<!-- @category: efficiency --><!-- @category: quality --><!-- @category: ux --><!-- @category: knowledge --><!-- @category: architecture -->For LIGHTWEIGHT sessions: Generate 0-2 recommendations max. For STANDARD sessions: Generate 2-5 recommendations.
Display recommendations grouped by scope (PROCESS vs PROJECT):
Present each recommendation with a number. User can approve all, selectively approve or deny in natural language. Denied items are noted as "considered but declined" in the retro file.
Route approved items to the most appropriate memory file (patterns.md, decisions/, glossary.md, product-context.md, or the retro's Improvement Backlog).
Create retrospective file at: .claude/memory/sessions/YYYY-MM-DD-retro.md
Format:
# Session Retrospective — YYYY-MM-DD
## Session Summary
[Brief 2-3 sentence summary of what happened in this session]
## Improvement Log
### Approved
- [Recommendation 1] → routed to patterns.md
<!-- @category: quality -->
- [Recommendation 2] → routed to decisions/ADR-NNN.md
<!-- @category: architecture -->
### Declined
- [Recommendation 3] — reason: [user's reason or "user declined"]
## Improvement Backlog
- [ ] [Future improvement idea 1 — actionable, specific]
- [ ] [Future improvement idea 2]
## Evidence
- Corrections: [N] detected, [M] processed
- Observations: [N] tool uses, [M] failures
- Friction signals: [list of friction patterns detected]
- Facets data: [available/unavailable] — include outcome, friction counts, and satisfaction summary if available
- Session depth: [LIGHTWEIGHT/STANDARD]
---
*Generated by /memory-reflect*
For LIGHTWEIGHT sessions, produce a minimal retro:
# Session Retrospective — YYYY-MM-DD
## Summary
[Short session with minimal activity. No notable improvements identified.]
## Evidence
- Observations: [N] tool uses
- Corrections: 0
---
*Generated by /memory-reflect (lightweight)*