learn-history

Analyze the full session history for this project to extract knowledge worth persisting.

Session logs location

Session logs are JSONL files at the project config path. Find them with:

ls -lhS ~/.claude/projects/$(echo "$PWD" | tr '/' '-')/*.jsonl

Each line is a JSON object with a type field (user for user messages, others for assistant responses). User messages have .message.content with the prompt text.

Process

1. Read the current state of AGENTS.md, .claude/agents/*.md, and the memory index for this project at ~/.claude/projects/$(echo "$PWD" | tr '/' '-')/memory/MEMORY.md (skip if absent)
2. List all session files, sorted by size (larger sessions have more content worth mining)
3. For each session, extract user messages and assistant text responses (skip tool results — they're too large). Focus on:
   • User corrections and redirections
   • Patterns that were discovered after expensive exploration
   • Knowledge that was needed repeatedly across sessions
4. Cross-reference findings against what's already documented to avoid duplicates
5. Group findings by destination:
   • AGENTS.md — brief conventions/rules (1-2 lines each)
   • .claude/agents/*.md — agent-specific knowledge
   • Memory — user preferences, project context, feedback
   • Nowhere — already documented or too specific to recur
6. Present a summary of all proposed changes for approval. Do NOT write anything until approved.

Guidelines

• Prioritize findings that recur across multiple sessions — those have the highest ROI
• Keep AGENTS.md entries concise. If something needs more than 2 lines, it probably doesn't belong there.