I got tired of repeating myself to every AI tool I use, so I built an open protocol to fix it

https://github.com/SMJAI/open-memory-protocol

I use Claude/Codex for coding, ChatGPT for drafting, Gemini for research, and Cursor inside my editor. They are all genuinely useful. But none of them know anything about each other.

Every time I switch tools I start from zero. I tell Claude my preferences. Then I tell ChatGPT the same preferences. Then Cursor. Then my terminal agent. It is the same few paragraphs of context, over and over, every single session.

That bothered me more than it should have. So I started thinking about why it happens.

The answer is obvious once you see it: there is no shared layer. Every AI company built their own memory system, inside their own walls, for their own product. Your memories in Claude belong to Claude. Your history in ChatGPT belongs to ChatGPT. None of it moves.

This is the same problem RSS solved for blog feeds in 2000. Not a technology problem. A standards problem.

So I built Open Memory Protocol.

OMP is an open specification for how AI tools store, retrieve and share memory. A reference server you can self-host or use for free on Railway. And a set of adapters that connect to the tools you already use.

Here is what it actually does today:

Claude Desktop remembers things automatically via MCP. You do not ask it to. It just does it.
The browser extension sits on Claude, ChatGPT, Gemini and Perplexity. When you open a new chat, it shows a toast: "Continue from ChatGPT? Here is what you were discussing." One click and the new AI picks up mid-conversation.
There is a mobile PWA you can add to your home screen. Open it anywhere, see your memories, generate a handoff brief and paste it into any AI app.
The server is live right now. No install needed to try it.
The handoff brief is the part people seem to find most useful. It is not a raw memory dump. It generates a natural paragraph explaining the thread of the conversation and what you were trying to figure out, shaped for the target model. You paste it in and the new AI responds as if it was already in the room.

What I have learned so far

The hardest part was not the code. It was resisting the urge to over-engineer the memory layer itself. A lot of people's first reaction is "but how do you decide what to remember?" and that is a real question. OMP uses AI extraction to filter what is worth keeping, and semantic search to surface only what is relevant to the current conversation. But the protocol is intentionally neutral on architecture — tools can implement their own context shaping on top of the shared store. The point is the shared store exists at all.

The second thing I learned is that people do not want to think about memory. They want their AI to just know them. The goal is to make OMP invisible. It runs in the background, saves quietly, and the AI is simply better next session. No manual steps.

Where it is now

The server is live and free to try at omp-server-production.up.railway.app. The repo is open source on GitHub at github.com/SMJAI/open-memory-protocol under Apache 2.0.

Early traction is coming from developers who use multiple AI tools daily and feel the context-switching pain. The Hacker News thread had some good pushback on the context shaping question which helped sharpen the thinking.

Next on the roadmap is semantic search with embeddings and per-project memory namespacing. Eventually multi-user support so teams can share a memory layer.

What I am looking for

Feedback on the protocol design itself — I want this to be community-owned, not something one person controls. If you build AI tools and want to add OMP support, the adapter interface is small and the docs are there.

And honestly, if this solves a problem you have, a star on GitHub goes a long way for visibility.

What does your current setup look like for carrying context between AI tools? Curious whether other people have hacked together their own solutions for this.