Every time you switch between Claude, ChatGPT, and Cursor, your context resets to zero. You spend the first five minutes re-explaining what you're working on, what you prefer, what's already been decided. It's 2026 and we're still treating AI memory like a walled garden. The Open Memory Protocol (OMP) wants to change that with an open specification for shared AI memory storage—and it's live now.

What OMP Actually Does

The core idea is straightforward: one self-hosted SQLite database that every AI tool reads from and writes to. Claude saves memories via MCP, ChatGPT talks through the browser extension, your custom agent hits the REST API directly—all reading from the same source of truth. The server runs in Docker on a single command (docker run -p 3456:3456 ghcr.io/smjai/omp-server), and you're off. No vendor lock-in, no cloud dependency, just a JSON-based memory object schema that defines episodic events, semantic facts about user preferences, and procedural how-to knowledge.

The Memory Object Schema

OMP defines a canonical format for memories—content, type, source tool, session ID, tags, timestamps, optional embeddings for semantic search. When Claude learns you prefer TypeScript over JavaScript, it saves a memory object to your local server. Switch to Cursor, and that same fact is available in the next query. The spec includes standard CRUD endpoints plus semantic search via /v1/memories/search. If you want pgvector support for embeddings at scale, that's on the v0.4 roadmap.

Current Adapter Status

The Claude MCP adapter (npx omp-mcp) is live and functional—point it at your local server with environment variables OMP_SERVER and OMP_API_KEY, add a system prompt to use omp_remember/omp_recall automatically, and Claude silently saves context without bothering you. The browser extension works with ChatGPT, Claude.ai, Gemini, and Perplexity—a floating brain icon lets you inject memories into any chat on the web. OpenAI Assistants, Cursor, Copilot, and Gemini adapters are marked "help wanted" in the docs, which means the spec is solid but adoption needs community muscle.

The Privacy Angle

OMP's creators make a point that resonates with anyone who's watched AI companies quietly expand their training data: your memories should live on hardware you control. The server defaults to self-hosted, there's no cloud offering forcing you in, and export/import is plain JSON for portability between servers or providers. Bearer token auth guards the API with per-tool keys. This isn't just open source philosophy—it's architecturally baked in.

Roadmap: v0.1 to v1.0

We're currently at v0.1 (core spec, reference server, MCP adapter). The roadmap shows AI-powered memory extraction and compression coming in v0.2, cross-model conversation handoff via the browser extension in v0.3, semantic search with pgvector in v0.4, per-project namespacing in v0.5, multi-user access control in v0.6, and a stable v1.0 spec submission to an open standards body eventually. It's early—but the foundation is there.

Key Takeaways

  • One self-hosted memory server replaces context-rebuilding across every AI tool you use
  • Claude MCP adapter works today; browser extension covers web-based ChatGPT and others
  • Open source under Apache 2.0, no vendor lock-in, plain JSON export for portability
  • Help wanted on adapters for Cursor, Copilot, Gemini—community contribution is the bottleneck

The Bottom Line

OMP solves a real problem that most users have accepted as normal. The memory silo situation is embarrassing in an era of cross-platform everything, and this spec deserves attention from anyone running multiple AI tools daily. Grab the Docker image, spin up a server this week, and see what it feels like when your AI actually remembers you.