claudemnesia

claude-in-claude — the cure for Claude's amnesia.

Claude Code's /compact summarizes a long conversation to save context — but that's lossy: it kills the exact wording, the dead-ends, the things you tried. The full, un-compacted conversation is still on disk in the .jsonl transcript.

claudemnesia recovers it. It extracts the real user↔Claude exchange — dropping tool noise, collapsing pasted log/code dumps so the vibe survives — and, over an MCP server, lets a running Claude search and recall its own un-compacted conversation, right after a /compact.

On a real 352 MB transcript: 352 MB → ~7 MB (98% smaller) — 4,176 messages, 243 dumps classified, 24 compaction markers found.

Install

uv tool install .        # everything: CLI + local hybrid search (no torch) + MCP server
claudemnesia --help

Local semantic search uses fastembed (ONNX, no PyTorch) — a base dependency, so the default install already has working hybrid search. '.[semantic-openai]' / '.[semantic-torch]' add the OpenAI / sentence-transformers embedders.

Quickstart

# Extract the latest conversation of the current project to Markdown:
claudemnesia --latest -f md -o conversation.md

# Recall: rank this conversation's turns against a question (cited, sub-second):
claudemnesia --latest --search "what did we decide about the dump threshold?"

# Orient after a /compact — the same surface the MCP tools expose:
claudemnesia --latest --recent 8            # the last 8 turns
claudemnesia --latest --outline             # one-line preview of every ask
claudemnesia --find-across "TODO_MARKER"    # exact search across all your sessions

→ Full reference: docs/cli.md

Recall for Claude itself

Expose claudemnesia over the Model Context Protocol so a running Claude can search its own transcript — the claude-in-claude idea. After a /compact, it recalls exactly what the summary dropped.

uv tool install .
claude mcp add --transport stdio claudemnesia -- claudemnesia-mcp

Everything defaults to the current conversation (via CLAUDE_CODE_SESSION_ID), so recall("what did we decide about X") just works post-compaction.

→ Tools, the wake-up flow, env vars: docs/mcp.md

What it does

• Extract — the real exchange, tool-noise dropped, dumps collapsed non-destructively (~98% smaller).
• Recall — per-turn search (BM25 + semantic, fused with RRF), cited, no vector DB.
• Navigate — recent / outline / since-compact, read a turn, read a hidden dump.
• Operate — list_files / list_actions: what the session changed, not just what it said.
• Across sessions — "have I solved this before, in another conversation?"

Documentation

• CLI — extract, the navigate/inspect verbs, every option, output formats.
• MCP server — the recall tools, the post-/compact wake-up flow, env vars.
• Recall & search — lexical / semantic / hybrid, the embedding cache, cross-conversation.
• Architecture & internals — signal vs. noise, the classification model, the hexagon, library use.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Changes are tracked in CHANGELOG.md.

License

MIT.