About

Memory should belong to the human, not the model.

mnueron is a persistent memory layer for AI tools. Built local-first, open-source, and provider-agnostic so your context follows you across Claude, Cursor, Windsurf, Cline, and anything else you use — without a vendor owning it.

The premise

Every AI conversation today is amnesiac.

You explain your project conventions to Claude. You decide on an architectural pattern with Cursor. You catalog a finicky API in ChatGPT. Tomorrow, you start over — each tool is a goldfish.

The fix is obvious once you say it out loud: memory should be a layer, not a feature. It should live next to the user, follow them between tools, and never be hostage to one vendor's subscription.

That's mnueron. A small Node engine that runs locally, exposes six MCP tools (save, recall, list, get, delete, threads), and auto-wires itself into every AI tool already on your machine. Use whichever LLM you want. Switch any time. Your context comes with you.

Stack at a glance

Real boring tech making one interesting thing.

Engine
Node 20+ • TypeScript • SQLite + FTS5 + sqlite-vec
Hosted backend
Postgres 15+ • pgvector • Express • RLS-enforced isolation
Auth
bcryptjs • httpOnly session cookies • optional SSO
Embeddings
Local Transformers.js (no API key) — pluggable
Client surfaces
MCP server • Chrome extension • Python + C# SDKs
Principles

Four lines we won't cross.

These aren't aspirations — they're design decisions baked into the architecture so we can't quietly walk them back later.

Local-first by default

Your memories live in ~/.mnueron/memories.db on your machine. We can't read them. There's no telemetry, no upload, no 'opt out' to remember to flip. The hosted backend exists for people who want cross-machine sync — and is itself open source.

Provider-agnostic, not provider-aligned

mnueron doesn't care if you use Claude, GPT-4, Gemini, Mistral, or Ollama. It's the memory layer underneath, not a wrapper around one vendor. The same memories surface in every tool you use.

Secrets never reach storage

Thirteen regex patterns plus Bearer/Basic-auth handling redact AWS, GitHub, OpenAI, Anthropic, Stripe, Slack, JWT, and PEM keys at write time. The original text is never persisted — not even to a temp file.

Open core, no rug-pulls

The client is MIT — fork it, ship it, sell on top of it. The hosted backend is FSL-1.1-Apache-2.0 — read it, contribute, self-host. Each version auto-converts to Apache 2.0 two years after release, so you can never be locked out of code you depend on.

Try it for the price of a `brew install`.

mnueron installs in under a minute and runs entirely offline. No account, no card. Sync is optional and shows up only when you ask for it.