DEV Community

Nucleus OS
Nucleus OS

Posted on

I got tired of losing Claude Code context between sessions, so I built a daemon

#go #sqlite #ai #claudecode

Every time I closed a Claude Code session, the conversation was gone.

Not archived. Not searchable. Just gone.

I'd start a new session and spend the first 10 minutes re-explaining context I'd just worked through. The "Summarize our last conversation" prompt only goes so far when you've had 300 sessions across 8 weeks.

So I built a fix.

What eideticd does

It's a Go binary that runs at login (launchd on macOS, systemd on Linux). It uses fsnotify to tail Claude Code session files in real time. Every message you type becomes an engram — a row in a local SQLite-WAL database — within <50ms of the file write.

Nothing leaves your machine.

# One-liner install
curl -fsSL https://eidetic.works/install.sh | sh

# Or Homebrew
brew tap eidetic-works/nucleus
brew install eideticd
Enter fullscreen mode Exit fullscreen mode

The numbers that matter

After 2 weeks of dogfooding on my own machine:

  • 141,502 engrams captured across sessions
  • P95 retrieval: 0.27ms on that live dataset
  • SLO was 100ms — we cleared it by 370×
  • Capture latency: <50ms on all measured writes

MCP bridge — let your AI query its own history

The companion Python package exposes the daemon's Unix socket API as MCP tools:

pip install eidetic-mcp
claude mcp add eidetic -- python -m eidetic_mcp.server
Enter fullscreen mode Exit fullscreen mode

After that, you can ask Claude Code:

"What was I debugging last Tuesday?"

And it pulls from your engrams in <1ms. No cloud, no API key, no subscription.

Architecture decisions worth knowing

Pure-Go SQLite via modernc.org/sqlite — not the CGO mattn/go-sqlite3. This matters for distribution: CGO + cross-compile silently strips SQLite and produces a binary that crashes at runtime. The pure-Go driver cross-compiles cleanly to darwin-arm64, linux-amd64, and windows-amd64 from a single host. I learned this the hard way.

WAL mode mandatory — PRAGMA journal_mode=WAL; synchronous=NORMAL is the only mode that gives concurrent readers without blocking the writer. Write-append-only shape + single-writer pool (SetMaxOpenConns=1) + separate read pool (8 conns) = zero "database is locked" errors under load.

FTS5 full-text index — AFTER INSERT/DELETE triggers maintain a full-text index on the content column. Boolean operators, phrase queries, relevance ranking — all the SQLite FTS5 goodness, sub-millisecond.

Single binary, no daemon manager required — eideticd -install writes the launchd plist and bootstraps it. Uninstall is eideticd -uninstall. No Docker, no Python runtime, no config files to manage.

What shipped in v0.0.25 (this week)

  • Compliance daemon (eideticd-compliance): reads a retention-policy.json, purges engrams older than the configured TTL per surface, writes an audit log, exits. Designed for cron/launchd/systemd timer.
  • PyPI package: pip install eidetic-mcp — the MCP bridge is now a proper PyPI package, not a pip install -e from source.
  • Homebrew formula: brew tap eidetic-works/nucleus && brew install eideticd

Try it 

curl -fsSL https://eidetic.works/install.sh | sh
Enter fullscreen mode Exit fullscreen mode

Source: github.com/eidetic-works/eidetic-daemon (MIT)

Happy to answer questions about the SQLite design, the fsnotify watcher, or the retention system.

Top comments (0)