Lamarck

lamarck

Automatically learns from your Claude Code conversation history and writes what it discovers into `MEMORY.md` files that get loaded into future sessions. Run it once, on a cron, or as a post-conversation hook — it tracks what it's already processed so it never re-scans the same sessions.

Why "Lamarck"?

Jean-Baptiste Lamarck proposed that organisms pass on traits acquired during their lifetime to the next generation. A blacksmith's children inherit strong arms. That's exactly what this tool does — traits Claude acquires during a coding session (patterns, pitfalls, preferences) get inherited by future sessions through MEMORY.md. Darwinian evolution says only random mutations survive; Lamarckian evolution says experience matters. Your Claude gets smarter because of what it lived through.

How it works

Claude Code sessions (.jsonl)
        │
        ▼
┌─────────────────────┐
│  Incremental scan   │  Reads only new lines from ~/.claude/history.jsonl
│  (byte-offset cursor)│  using inode + offset tracking
└──────────┬──────────┘
           ▼
┌─────────────────────┐
│  Per-project queue   │  Discovered sessions queued for processing
│  with retry tracking │  Failed sessions retry on next run
└──────────┬──────────┘
           ▼
┌─────────────────────┐
│  LLM reflection      │  Claude extracts diary entries from transcripts:
│  (diary + deltas)    │  accomplishments, decisions, preferences, pitfalls
└──────────┬──────────┘
           ▼
┌─────────────────────┐
│  Quality gates       │  Similarity dedup, content filtering,
│                      │  category quotas, minimum scores
└──────────┬──────────┘
           ▼
┌─────────────────────┐
│  Playbook curation   │  Add, merge, deprecate, promote/demote bullets
│                      │  with conflict detection
└──────────┬──────────┘
           ▼
┌─────────────────────┐
│  MEMORY.md export    │  Writes to each project's mem