claude-find is used to extract deep memory across Claude Code sessions and retrieve it whenever you need it.
Purpose
- Perform semantic search across all past Claude Code sessions, finding context by meaning and keywords.
- Retrieve raw conversation transcripts rather than compressed summaries, so Claude gets complete information: reasoning process, constraints, failed attempts, and decisions made.
Key capabilities
- Index all Claude Code session JSONL files under ~/.claude/projects/.
- Extract user and assistant messages, compact summaries, and file paths from tool calls.
- Enrich each chunk with metadata context (project, branch, files, date) to improve retrieval effectiveness.
- Embed conversation chunks with qwen3-embedding via Ollama (supports GPU acceleration).
- Use hybrid semantic and keyword (FTS5) search, and merge results via Reciprocal Rank Fusion.
- Return raw conversation chunks so Claude can perform comprehensive analysis with full context.
How to use
- First install Bun and Ollama.
- Run bunx claude-find setup. This command detects your platform and guides you through filling in missing dependencies.
- setup starts Ollama, pulls the embedding model, sets session retention to permanent, and registers the MCP server with Claude Code.
- Sessions are indexed in the background at startup. Search is available immediately and returns progressively more complete results as indexing continues.
- In any Claude Code session, Claude performs semantic search across your past sessions, finds relevant conversations, and synthesizes context: what was tried, what failed, what constraints you set, and what decisions were made.
- After upgrading, run bunx claude-find index to rebuild the index with the latest improvements.
Features and benefits
- Search raw transcripts: no information is lost due to compression.
- Retroactive: works immediately for all existing sessions, with no hooks required.
- Permanent history: setup disables Claude Code's 30-day session cleanup, making your sessions searchable permanently.
- Non-blocking: indexing is done in the background at startup, and search works instantly even while indexing is in progress.
- Uses compact summaries: that is, Claude's own session understanding, and boosts its weight in ranking.
- Indexes tool call metadata: you can search by files involved and errors encountered.
- Fast: Ollama plus GPU keeps indexing fast and memory manageable.
Applicable scenarios
- When you need to review approaches, constraints, and decisions discussed in past Claude Code sessions.
- When you want to find past conversations involving specific files or errors.
Limitations and dependencies
- Depends on Ollama (models are downloaded automatically on first use).
- Requires Bun and Ollama to be installed.
- The provided material does not include information such as platform support scope, fees, or pricing.
