learn-session

Review the current conversation to extract knowledge worth persisting for future sessions.

What to look for

1. User corrections — where the user redirected the approach, rejected a suggestion, or clarified a preference
2. Discovered patterns — implementation patterns, conventions, or architectural knowledge that was hard to find and would save time next session
3. Agent/tooling gaps — things the tester or code-reviewer agents should know but don't

What NOT to persist

• Things already documented in AGENTS.md or .claude/agents/
• Code-level details derivable by reading the source
• One-off debugging context that won't recur
• Verbose implementation guides — keep entries brief (1-2 lines for AGENTS.md)

Process

1. Read the current state of AGENTS.md and .claude/agents/*.md
2. Read the memory index for this project at ~/.claude/projects/$(echo "$PWD" | tr '/' '-')/memory/MEMORY.md (skip if absent)
3. Review the conversation so far: what was the task, what was learned, what corrections were made
4. For each finding, determine where it belongs:
   • AGENTS.md — brief convention or rule that applies to all sessions (1-2 lines max)
   • .claude/agents/*.md — specific to the tester or code-reviewer agent
   • Memory — user preferences, project context, feedback (not code patterns)
   • Nowhere — already documented or too specific to recur
5. Present the proposed changes for approval. Do NOT write anything until approved.