Session Learning Extractor

Scans the current session for patterns, corrections, and feedback, then proposes updates to memory and config.

Usage

• /learn — Scan session, propose updates
• /learn <topic> — Focus extraction on a specific topic

Workflow

Phase 1: Scan Session

Review the conversation for:

1. PR review feedback — corrections, style preferences, pattern violations flagged
2. Errors encountered — what went wrong, what fixed it, root cause
3. Patterns discovered — new conventions, architectural decisions, workflow preferences
4. Corrections — things the user corrected (naming, approach, style)
5. Explicit instructions — "always do X", "never do Y", "remember this"

Phase 2: Compare Against Config

Check each learning against existing config layers:

Layer	Path	Purpose
CLAUDE.md	./CLAUDE.md	Project conventions (checked in)
Rules	.claude/rules/*.md	Contextual rules (loaded by glob/description)
Agents	.claude/agents/*.md	Agent personalities and strategies
Skills	.claude/skills/*/SKILL.md	Skill definitions
Memory	~/.claude/projects/.../memory/	Persistent cross-session memory

Phase 3: Identify Gaps

For each learning, classify:

• Missing — not captured anywhere, needs to be added
• Contradicts — conflicts with existing rule/memory
• Vague — existing rule is too general, needs sharpening
• Wrong layer — captured but in the wrong place (e.g., session-specific in CLAUDE.md)
• Already covered — no action needed

Phase 4: Propose Updates

Present findings grouped by target file:

## Session Learnings

### New Patterns Found
1. [pattern]: [evidence from session]

### Proposed Updates

#### memory/MEMORY.md
- Add: [new entry]
- Update: [existing entry] → [revised entry]

#### .claude/rules/elixir.md
- Add: [new rule with evidence]

#### CLAUDE.md
- Update: [section] — [what changed and why]

### No Action Needed
- [learning already covered by existing config]

Rules

• Evidence-based — every proposed update must reference a specific moment in the session
• Right layer — session-specific → memory, stable patterns → rules, project conventions → CLAUDE.md
• No duplicates — check existing content before proposing additions
• Worktree-aware — always write to main tree's memory path, not worktree
• Confirm before writing — present proposals, wait for user approval before modifying any file
• Concise memory — MEMORY.md must stay under 200 lines (truncated beyond that)