claudegrep

grep, but semantic, for your Claude Code conversation history. Local-first, private, fast.

Every Claude Code session vanishes the moment you close it. claudegrep indexes your ~/.claude/projects/ JSONL files into a single SQLite database and gives you a fast CLI (recall) to find any past turn — hybrid BM25 + vector search with optional cross-encoder reranking.

$ recall search "sqlite vec hybrid"
1. 2026-04-29  claude-recall  score=-18.4
   how should we store the chunk vectors — sqlite-vec or lancedb?
   [assistant] for <10M vectors sqlite-vec wins on ops simplicity…

2. 2026-04-22  gstack         score=-14.1
   …

Status

Week 5 / 6 — recall serve is live: a local FastAPI + HTMX web UI at http://127.0.0.1:7777 with hybrid search, inline filters, session viewer, and one-click "copy turn" to clipboard. v0.3.0 ships it. See PLAN.md.

Install

# CLI only (~120 MB with embedder model)
uv tool install claudegrep
# or:
pipx install claudegrep

# CLI + local web UI
uv tool install 'claudegrep[serve]'

# first run
recall index
recall search "your query"
recall serve              # opens http://127.0.0.1:7777

Dev install:

git clone https://github.com/lbbstarry/claudegrep
cd claudegrep
uv sync
uv run recall index

Commands

Coming in Week 6+:

• bge-m3 + query expansion (planned biggest single eval bump)
• Pro tier: encrypted cross-machine sync, voyage-3-lite embedder
• v1.0 + Product Hunt launch

Architecture

~/.claude/projects/*.jsonl
    │
    ▼
parsers/claude_code.py    typed Message records
    │
    ▼
ingest/chunker.py         per-turn chunks (one user msg + following assistant)
    │
    ▼
embed/local.py            sentence-transformers (bge-small-zh) + disk cache
    │
    ▼
store/                    SQLite + FTS5 + sqlite-vec (vec0 virtual table)
    │
    ▼
search/                   bm25 · vector (KNN) · hybrid (RRF) · rerank (cross-encoder)
    │
    ▼
cli.py                    Typer app

Why per-turn chunks? Per-message loses Q-A pairing; sliding-window inflates the index 3-5×. A turn averages ~1.2k tokens — perfect for bge-m3's 8k context, no truncation needed.

Why SQLite + FTS5 + (later) sqlite-vec? Single file, zero ops, ships with the wheel, hybrid search is a JOIN away. Beats Chroma at this scale.

Eval

51 hand-labeled (query, relevant_chunk_id) pairs from real developer sessions in tests/fixtures/queries.jsonl. Index size: 6,593 chunks across 127 sessions / 6 projects. Reproduce with python -m claude_recall.eval.run.

Hybrid + rerank gives +118% Recall@10 and +95% nDCG@10 over BM25. Reranker latency is dominated by CPU cross-encoder inference; GPU or bge-reranker-v2-m3-onnx will reduce it. Numbers are CPU-only on a WSL2 Ryzen laptop.

Why the absolute numbers look modest: the eval queries are deliberately short (median 5 chars) and developer-domain-specific, e.g. prefab, figma示例. That is the realistic distribution for "I vaguely remember talking about this last month" — and the gap between methods, not the absolute floor, is what matters. Query expansion and bge-m3 (8k context, multilingual) are next.

Roadmap

• Week 1 — Typer CLI, SQLite + FTS5, incremental ingest, BM25 search
• Week 2 — bge-small-zh embeddings, sqlite-vec, RRF hybrid, reranker, 51-query eval
• Week 3 — show/inject/export, watch daemon, filter DSL, --rerank flag
• Week 4 — v0.2 on PyPI, eval blog post drafts (EN + ZH)
• Week 5 — recall serve (FastAPI + HTMX local UI), v0.3
• Week 6 — v1.0, Pro tier (cloud sync, Voyage embedder), Product Hunt

Privacy

Everything stays on your machine. The index is a single SQLite file under your OS's user data dir (~/.local/share/claude-recall/recall.db on Linux). No network calls are made by the OSS build.

License

Apache-2.0