memo-retro

Analyze recent checkpoints to surface learnings and suggest memory updates.

Run periodically (weekly or after major milestones) to close the learning loop.

Steps

1. Detect the project from the current working directory. Construct the project slug by replacing all / with - in the full path.

2. Call the list_checkpoints MCP tool with limit: 10 (no date) to fetch the 10 most recent checkpoints.

3. Load current memory state:

   • Read all files in ~/.claude/projects/[PROJECT_SLUG]/memory/
   • Read MEMORY.md in the memory directory
   • Read the project's CLAUDE.md (in the working directory root)

4. Check git history for correction patterns:

   git log --oneline -30
   

   Look for reverts, quick-follow fix commits (e.g., "fix: ..." immediately after a feature commit), and patterns indicating repeated mistakes.

5. Analyze the checkpoint data:

   Extract learnings:

   • Pull the learnings array from each checkpoint (may be empty or absent on older checkpoints)
   • Also extract signals from decisions_made, blockers, and what_was_built

   Frequency analysis:

   • Count how many times each learning (or semantically similar learning) appears
   • Single occurrence = note only
   • 2+ occurrences = high confidence, propose as rule

   Cross-reference:

   • Is this learning already captured in a memory file? → skip
   • Is this learning already a CLAUDE.md rule? → skip
   • Is this a correction/preference? → propose as feedback memory
   • Is this about the project? → propose as project memory

6. Present the report:

## 🔄 Retro — [PROJECT_NAME] — [DATE]

### 🧠 Memory Health

- 📁 Total memories: [count of files in memory dir]
- 📏 MEMORY.md line count: [N]/200 (⚠️ warn if >150)
- 🪦 Stale memories: [any referencing completed/outdated work]
- 🔀 Duplicate/overlapping memories: [any found]

### 📚 Learnings Summary

**🔁 Recurring (2+ times) — propose as rules:**

- [learning] — appeared in [N] checkpoints
- ...

**📌 Single occurrence — note only:**

- [learning]
- ...

### 🔍 Git Correction Patterns

[Reverts, quick-follow fixes that indicate recurring mistakes, or "✅ None detected"]

### 🛠️ Proposed Changes

**📝 Memory file actions:**

- ➕ **Create:** [new memory file if a pattern warrants it, with suggested filename and content]
- ✏️ **Update:** [existing memory with outdated info]
- 🔗 **Merge:** [overlapping memories to consolidate]
- 🗑️ **Prune:** [stale memories to delete]

**📐 CLAUDE.md rules** (only for recurring learnings not already captured):

1. Add to section [X]: "[exact rule wording]"
   _Reason: appeared in [N] checkpoints, e.g., [quote from learning]_

1. Ask: "Which changes should I apply? (all / list numbers / none)"

   Do NOT modify CLAUDE.md, memory files, or any project files without explicit approval.