claude-recall

Automatic, cheap, precise query of your Claude Code session archive — wired into the prompt flow via hooks so you never have to remember to search.

Status: beta

claude-recall is in daily production use against a 26,000+ message session archive across 21 projects. The API and hook contract are stable through v0.6.0. Known issues are tracked in GitHub (see Known Issues).

Install via PyPI: pip install 'claude-recall[embeddings]'. Beta = rough edges get filed, triaged, and shipped quickly. Recent cadence: seventeen issues filed (fourteen closed) across April 21–26, with three same-day patch releases on April 25 alone (a Claude Code v2.1.118 hook-schema regression, an embeddings-coverage honesty fix, and a v0.6.0 architectural fix that stops routine indexing from wiping vectors). Full history in the changelog.

Feedback welcome via GitHub issues. The tool handles single-user, single-machine session archives. Multi-user / shared-recall is explicitly out of scope for beta.

What's landed through v0.6.x: MVP + keyword extraction + auto project scoping + config-driven hook flags + stale-hook detection + opt-in semantic retrieval via Ollama + NativeAOT C# hook binary + init-hooks --force overwrite semantics + PyPI publish + honest embeddings-coverage reporting + content-hash-aware indexer (vectors no longer cascade-wipe on routine re-ingest). See the original plan, feature plans, and the changelog for the full history.

What it does

Claude Code writes every session to ~/.claude/projects/<project-slug>/<session-id>.jsonl — raw turn-by-turn log. That archive is already a goldmine; the problem is that after context compaction, Claude loses mid-session detail, and no native Claude Code feature surfaces prior-session context back into the active conversation. Manual grep works but rarely happens.

claude-recall closes that gap:

1. Indexes the .jsonl archive into a local SQLite database with FTS5 full-text search. Incremental, read-only against the archive.
2. Queries via CLI — ranked matches with snippets, sub-millisecond FTS5 lookup over hundreds of thousands of messages.
3. Injects relevant prior-session context into the active Claude Code session automatically via SessionStart and UserPromptSubmit hooks. You don't have to remember to look.

Install

pip install 'claude-recall[embeddings]'
claude-recall init-hooks
claude-recall index

Pip picks the right wheel for your platform:

• Windows x64 gets the NativeAOT hook binary bundled automatically (fast-path hook)
• macOS / Linux get the pure-Python shell-hook fallback

To turn on semantic rerank (optional but recommended):

ollama pull nomic-embed-text
# Edit ~/.config/claude-recall/config.toml (or %APPDATA%\claude-recall\config.toml):
#   [embeddings]
#   enabled = true
claude-recall embed --probe    # sanity-check Ollama path
claude-recall embed --verbose  # one-time corpus embed

The init-hooks output also points at the section of the INTEGRATION-GUIDE that walks through this.

Don't use pip install git+https://... — that builds from source and the NativeAOT hook binary is CI-built package data, not part of the source tree. A git+https install silently gives you the slower Python hook path on every platform.

Quickstart query

claude-recall search "regex patterns" --days 30 --limit 5
claude-recall search "something you roughly remember" --semantic    # requires embeddings enabled
claude-recall show <session-id>
claude-recall status

Architecture

┌─────────────────────────────────────┐
│  ~/.claude/projects/<project>/*.jsonl│
│  (Claude Code's raw session archive) │
└────────────┬────────────────────────┘
             │ read-only
             ▼
┌─────────────────────────────────────┐
│  claude-recall indexer               │
│  walks jsonl, extracts message pairs │
└────────────┬────────────────────────┘
             │ writes
             ▼
┌─────────────────────────────────────┐
│  SQLite DB with FTS5 virtual table   │
│  ~/.config/claude-recall/index.db    │
└────────────┬────────────────────────┘
             │ queried by
             ▼
┌──────────────────┐  ┌─────────────────────┐
│  CLI (manual)    │  │  Hooks (automatic)  │
│  recall search   │  │  SessionStart       │
│  recall show     │  │  UserPromptSubmit   │
└──────────────────┘  └─────────┬───────────┘
                                │ additionalContext
                                ▼
                      ┌─────────────────┐
                      │  Claude Code    │
                      │  active session │
                      └─────────────────┘

Hook latency expectations

Concrete numbers from dogfooding so you can calibrate. Latency scales with archive size:

The cold-model case is Ollama unloading nomic-embed-text between invocations (default keep_alive=5m). The first prompt after idle pays the model-load cost; subsequent prompts land on the warm model. If latency bothers you, tune [embeddings].request_timeout_seconds and consider keeping Ollama warm via another process.

If the semantic hook feels slow on your archive size, flip [embeddings].use_in_hook = false to run FTS5-only. You still get the --semantic flag for manual claude-recall search invocations where extra latency is acceptable.

Why this exists

Claude Code has excellent primitives — the .jsonl session archive, hooks with additionalContext injection, skills, auto-memory markdown files. It does not have native session-archive search as of April 2026. claude-recall fills that gap with a small, self-contained Python CLI plus two hook scripts, following the pattern established by Microsoft's auto-memory project for Copilot CLI — adapted for Claude Code's persistence model.

Known Issues

Live list of issues caught in beta use. Filing new issues is explicitly welcome — the beta is the feedback channel.

• Open issues: see github.com/LearnedGeek/claude-recall/issues for the current list.
• Recently closed (not a full changelog — see the CHANGELOG for that):
  • #16 — Routine re-indexing was silently cascade-wiping embedding vectors via FK ON DELETE CASCADE; status reported embeddings_ready: True while search returned "no vectors in index" for the same DB. Fixed architecturally in v0.6.0 with content-hash-aware indexer + non-destructive schema migration that preserves existing vectors across the upgrade.
  • #15 — Claude Code v2.1.118's strict-validation pass rejected the legacy flat hook shape claude-recall had been generating since v0.4, silently disabling all of the host project's settings as a side effect. Fixed in v0.5.4 with the schema-correct nested shape and shell: "powershell" for .ps1 hooks on Windows.
  • #11, #12 — Hook-latency claim on PyPI was true on small fixtures but wrong at 25k-message scale; binary --version was hardcoded constant rather than reading from the package. Both fixed in v0.5.2.
  • #8 — embed batches used to fail the whole 32-message batch when one input was oversized, silently dropping 27% of vectors. Fixed in v0.4.3 with per-input truncation, error-body surfacing, and per-message singleton fallback. Dogfooding validated 99.95% coverage on a 25k-message archive after the fix.
  • #4 — init-hooks --force merged with existing settings.json instead of overwriting, producing duplicate UserPromptSubmit entries when upgrading from a manual wiring. Fixed in v0.4.2.
  • #3 — init-hooks --force crashed on Windows native-binary wheels (missing shell-script sources in the wheel). Fixed in v0.4.1.
  • #2 — Hook didn't auto-scope to current project + hardcoded --days 30 ignored config. Fixed in v0.2.1.
  • #1 — UserPromptSubmit hook returned empty on natural-language prompts. Fixed via stdlib keyword extraction in v0.2.0.

Beta expectations: I find issues by using the tool. You may find different issues by using it differently. That's the point of a beta. File and describe — triage is fast.

Documentation

• PLAN.md — full implementation plan, specifications, acceptance criteria.
• ARCHITECTURE.md — deeper architectural design, rationale, alternatives considered.
• INTEGRATION-GUIDE.md — how to wire claude-recall into any Claude Code project.
• EMBEDDINGS-PLAN.md — v0.3 semantic-retrieval feature plan.
• HOOK-BINARY-PLAN.md — v0.4 NativeAOT C# hook binary plan.
• CHANGELOG.md — full release history.

License

MIT. See LICENSE.

Author

Mark McArthey / Learned Geek