Lossless AI Memory with Longhand for Claude Code

Описание и применение

Longhand — это локальное решение для безпотерьного хранения истории сессий Claude Code. Вместо суммаризации ИИ, оно индексирует JSONL-логи в SQLite и ChromaDB, обеспечивая быстрый семантический поиск, конфиденциальность и работу офлайн.

Задача

AI memory tools typically summarize sessions, resulting in lossy data where specific debugging steps, reasoning traces, and failed attempts are discarded. Additionally, relying on vendor context windows creates privacy risks, vendor lock-in, and dependency on network connectivity.

Результат

Longhand provides a lossless, offline, and private alternative to AI memory summarization. It indexes Claude Code's JSONL logs into local SQLite and ChromaDB stores, enabling fast semantic recall (~126ms) without API calls or data loss. It supports forensic replay of code states and is portable across models.

Рабочий процесс

• Install Longhand via pip: pip install longhand
• Run setup to backfill existing Claude Code history and install auto-ingest hooks: longhand setup
• Longhand automatically captures new sessions via a SessionEnd hook in Claude Code.
• Query history using the MCP server tools, e.g., longhand recall "that webhook fix from last week"
• Use specific tools like replay_file to reconstruct file states or get_session_timeline for chronological replays.

Доказательства

Semantic recall across 100+ sessions: ~126ms

Источник

Storage footprint: ~1GB for a heavy power user, 200–400MB typical

Источник

API calls per query: zero

Источник