"Open Memory Protocol: One Memory Store for All Your AI Tools"

10:00 pm

If you've ever told Claude about your coding preferences, switched to Cursor, and had to explain everything from scratch — the Open Memory Protocol (OMP) is for you.

OMP is a vendor-neutral specification that lets AI tools share memory about users and their context. Think of it as a single, self-hosted brain that every AI you use can read from and write to.

The Problem It Solves

Right now, every AI tool has its own siloed memory:

  • Claude knows what you told it yesterday. Cursor doesn't.
  • ChatGPT learned your preferences. Your custom agent hasn't.
  • Copilot saw your code style. Your terminal AI is starting from zero.

Every time you switch tools, your AI forgets you. OMP aims to fix this with an open protocol, the same way SMTP unified email and HTTP unified the web.

How It Works

The architecture is refreshingly straightforward:

  1. You run a self-hosted OMP server — it's a Node.js server with SQLite or PgVec storage, deployable in one command (npx omp-server or Docker)
  2. Each AI tool connects via an adapter — MCP for Claude Desktop, SDKs for Cursor, REST API for custom agents
  3. Memory is stored as structured objects with content, type flags (semantic, episodic, procedural), tags, and source metadata

The server exposes CRUD endpoints for memories, plus search and a conversation handoff API that generates natural language briefs when switching between tools.

The Bridge Extension

The most clever piece is the OMP Bridge browser extension. It silently saves your ChatGPT conversations to your OMP server every two minutes. When you open Claude.ai, a toast notification asks: "Continue from ChatGPT? [topic]" — click it, and OMP injects a natural handoff brief. Claude picks up exactly where you left off, even though it wasn't in the original conversation.

Self-Hosted and Open

OMP is Apache 2.0 licensed. Your memory data stays on your own server — no third party involved. The spec defines memory objects, storage format, and the HTTP API, with TypeScript and Python SDKs for building OMP-compatible tools.

It's early days (the spec is at v0.2), but the approach is right. AI memory shouldn't be a competitive moat for platform vendors — it should be a portable layer that follows you between tools. OMP is one of the more practical attempts I've seen at making that happen.