Nothing is summarized and no model decides what was worth keeping, so what you get back months later is what actually happened — and it survives the source files being deleted. Output is capped by design: a full recall across 100+ sessions comes back around 4K tokens, far less than reading a single raw session file.
A local memory store for Claude Code. It ingests session transcripts into SQLite — every tool call, file edit and thinking block, stored verbatim — and makes them searchable, replayable and recallable in natural language.
- Recall — fuzzy natural-language recall across every session, the recommended first call
- recall_project_status — a git-aware summary of where a project was left, with commits, issues and last outcome
- Search — filtered search by session, project, tool, file and event type, optionally with surrounding conversation
- list_sessions and get_session_timeline, with tail pagination for "how did it end"
- find_episodes — structured problem-to-fix pairs, with the referenced events, diff and post-fix file state
- replay_file and get_file_history — reconstruct a file at a point in time, or every edit across sessions
- find_commits, list_projects, list_plans and get_stats
- Reconcile — re-ingest and re-attribute from inside a session; it defaults to a dry run
Pip install longhand, then longhand setup to ingest your history, install hooks and configure MCP. Python 3.10+, version 1.0.1. It reads your local session files and makes no API calls; longhand demo runs a sandboxed walkthrough that never touches them.
One command — pip install longhand