Non-negotiable rules:

1. Only record learnings that belong in agent memory.
2. Keep one central learnings source of truth; do not invent parallel central files.
3. Sync approved learnings into every relevant agent surface that exists for the supported CLIs.
4. Do not propagate "Claude Code Only" learnings into engineer/reviewer agent prompts.
5. Get explicit user confirmation before modifying the learnings file or any agent file.

Update Agent Learnings

Inputs

• $request: Optional learning candidate, scope hint, agent name, or reminder about what the session revealed

Goal