The problem

Every AI coding session starts the same: paste the README, paste the architecture doc, paste the last error message, re-explain where you left off. If you switch between Claude Code, Cursor, and Gemini in the same project, every tool starts from zero.

nodestone fixes that. It's a project-level memory engine for AI-assisted development, and it's free.

What it does

nodestone tracks four things across sessions:

• Decisions — why you chose SQLite over Postgres, or that weird middleware pattern
• Tasks — what's done, what's next, with dependencies and critical path
• Drift — when you start implementing something different from what you planned
• Context — a compressed ~500 token pack that any AI can restore in one command

Install

pip install nodestone

That's it. Python 3.10+.

Quick start

1. Initialize in your project

nodestone init

Creates a .nodestone/ directory in your project root.

2. Save your first decision

nodestone decision add "Use SQLite instead of Postgres" --context "Only 3 users, no need for a server" --impact "Simpler deployment, no Docker needed"

3. Add a task

nodestone task add "Add user authentication" --depends "Set up database schema" --effort 4h

4. Snapshot context for the next session

nodestone pack

Prints a ~500 token context block. Copy it, paste it at the start of your next AI session. Done.

5. Restore on the other side

nodestone restore <fingerprint>

The other agent picks up exactly where you left off — decisions, tasks, file state, everything.

Real example: cross-agent handoff

# Session 1 — Claude Code
$ nodestone decision add "Use FastAPI over Flask" --context "Need async support for WebSockets"
$ nodestone task add "Build WebSocket handler" --effort 3h
$ nodestone pack
> Context fingerprint: ns_X7k2m9 (copied to clipboard)

# Session 2 — Cursor (next day)
$ nodestone restore ns_X7k2m9
> Restored: 3 decisions, 4 tasks, 2 file changes

No pastebin, no markdown file, no "as I mentioned yesterday...".

A note on trade-offs

• Works best with pack + restore flow. Real-time sync across agents is on the roadmap but not here yet.
• The scheduler (OR-Tools CP-SAT) is powerful but overkill for hobby projects — you can ignore it entirely.
• It's new. Documentation is good but the community is small. That said, the core loop (decisions + tasks + pack/restore) is solid.

What's next

• Fingerprint handoff (done): zero-loss context transfer
• Drift detection (done): alerts when implementation diverges from plan
• Plugins/triggers (done): webhooks, MCP tools, CLI hooks on state changes
• Real-time sync (next): multiple agents sharing context live

Links

• GitHub: https://github.com/massiron/nodestone
• Site: https://deepstrain.dev
• Install: pip install nodestone

If you're switching between AI tools on the same project, give it a try. Feedback and PRs welcome.