🚀 v0.4.0 — Claude Code Support is Here (+ a Sidecar That Just Works)

[dezgit2025]

🚀 v0.4.0 — Claude Code Support is Here (+ a Sidecar That Just Works)

auto-memory now speaks Claude Code.

When I started this project, the goal was simple: give AI coding agents memory that doesn't cost 200K tokens per session. We nailed that for GitHub Copilot CLI — 50 tokens, instant recall, zero deps.

The people wanted Claude Code! So i listened. Every time I'd start a new Claude session, it had zero idea what I'd been working on five minutes ago. Same amnesia, different agent. Also my experience with Claude Code were more inconssitent so your often spent compacting, /clear,etc so this should help with debugging.

What shipped in v0.4.0

A brand-new Claude Code provider — inspired by two community contributions: PR #8 from @osamarehman, which laid out a working JSONL parser and FTS5 schema, and PR #12 from @danielsimonjr, which proposed a clean file-backed provider integration. We used both as reference points, then went our own direction — building a separate binary with its own FTS5 index, security-hardening the parser, and adding our own provider/CLI/sidecar layer. The end result is its own thing — but those PRs were the spark that got us moving.

The big design decisions

1. Separate binary — session-recall-cc is a completely separate process from session-recall. A bug in the Claude Code provider can never crash your Copilot CLI workflow. They share zero code paths at runtime.

2. Agent-driven, not hook-driven — PR feat: Claude Code backend, Cursor/Aider backends, MCP server, export/prune commands #8 originally injected via Claude Code's SessionStart hook, mutating ~/.claude/settings.json. We rejected that. Instead, you add a one-line instruction to your CLAUDE.md and the agent decides when to call recall. Same philosophy as Copilot CLI — the agent has agency, not a hidden hook.

3. FTS5 with teeth — Not just basic full-text search. We added Porter stemming, unicode61 tokenization, bm25 column weighting (user messages ranked 5× higher than assistant output), prefix queries, and snippet extraction with ⟦highlight⟧ markers. Search actually finds what you're looking for.

4. Env var gate — SESSION_RECALL_ENABLE_CLAUDE_BACKEND=1 must be explicitly set. If you're a Copilot-only user, you'll never even know Claude Code support exists. Zero cost, zero noise.

What it looks like in practice

You: "What work was done on the API integration?"
Claude: *checks session-recall-cc* → finds 3 sessions from last week
Claude: "Based on your recent sessions, you refactored the auth middleware
         on April 23 and worked on database migrations on April 4..."

Instead of guessing or asking you to re-explain, the agent just knows.

session-recall-cc list --limit 5

ID	Repo	Summary	Last Seen	Turns
3b89ff09	myorg/backend-api	API integration work	2026-05-02	1
7e63b4b4	myorg/infra	Where are Claude Code session files?	2026-05-02	1
78c36e1f	myorg/infra	How do I update Claude Code?	2026-05-02	2

session-recall-cc search "database migration"

ID	Repo	Snippet
bc112c14	myorg/backend-api	⟦Database⟧ ⟦migration⟧ for user table — added indexes on email and created_at
01bf810e	myorg/infra	Check if staging ⟦database⟧ has pending ⟦migration⟧ scripts

The sidecar — tested on my Mac (Apple Silicon)

Here's something I'm personally excited about. The CLI does on-demand indexing — every time you call session-recall-cc, it checks if any JSONL files changed since the last index and incrementally updates. Typical latency: ~50ms.

But I wanted to go further. So we built a sidecar — a lightweight background process that pre-warms the FTS5 index every 5 minutes. On my M-series Mac, I set it up with launchd:

bash scripts/install-claude-sidecar.sh --install

That's it. One command. It auto-detected macOS, wrote a launchd plist, and now the index is always warm. First call after waking my laptop? Instant. No cold start.

The sidecar is completely optional — the CLI works great without it. But if you're the kind of person who wants sub-100ms on every single call (hi, that's me), it's a one-liner to set up on macOS, Linux, or WSL2.

How it works under the hood:

launchd fires every 5 minutes
  → python -m session_recall.providers.claude_code.sidecar --once
  → scans ~/.claude/projects/*/*.jsonl (mtime check — skips unchanged files)
  → upserts into ~/.claude/.sr-index.db (single SQLite transaction)
  → exits cleanly — no daemon, no memory footprint

Typical run on Apple Silicon: <50ms for incremental updates (only re-parses changed files). Full rebuild of 30 sessions: ~200ms.

The numbers

Metric	Value
New source files	7
New test files	7 (71 tests)
Total tests passing	268
Lines added	+4,174 (mostly tests + docs)
Security findings addressed	13 (from adversarial review of PR #8 design)
Runtime dependencies added	0 (still stdlib-only)
User config files mutated	0

Try it

pip install auto-memory[claude]
export SESSION_RECALL_ENABLE_CLAUDE_BACKEND=1
session-recall-cc health

Then add the instruction block to your CLAUDE.md — or ask your agent to read deploy/install-claude-code.md and it'll walk you through it with confirmation at every step.

What's next

• Fuzzy/semantic search (Phase 2) — trigram fallback + optional vector embeddings via sqlite-vec
• More providers — Cursor, Aider, Codex CLI (the architecture is ready)
• PyPI name reservations — auto-memory-cc, auto-memory-codex are claimed

Thanks to @osamarehman for PR #8 and @danielsimonjr for PR #12 — both surfaced Claude Code support as a real community need. PR #8's JSONL parsing approach and FTS5 schema, and PR #12's clean file-backed provider integration, were the reference points we built on.

---

auto-memory v0.4.0 is live on PyPI. Star the repo if this is useful — it helps others find it.

[osamarehman]

Thanks, PyPi Package still not updated with Claude support. @dezgit2025