How to Give Claude, Cursor, and ChatGPT the Same Memory

Why every new AI feels like a new hire

You explain your project to Claude. Next week you open Cursor and explain it again. Then ChatGPT, from scratch, one more time. Each tool keeps its own memory, and each one keeps that memory to itself. So you are not building knowledge, you are repeating yourself in three places and hoping the version in your head stays consistent across all of them. That is the real cost, and it sits underneath almost every complaint about AI forgetting who you are.

Built-in memory does not solve this, because built-in memory is one copy per vendor. What Claude learns about you stays with Claude. What ChatGPT learns stays with ChatGPT. There is no shared middle.

What one brain actually means

The fix is boring in the best way. You keep one folder of plain Markdown files on your own computer. Markdown is just text with light formatting, the kind of file you can open in any text editor and read without special software. That folder is the memory. Every AI you use connects to that same folder, reads from it, and writes back to it.

The connection happens over MCP, the Model Context Protocol, which is the open standard AI assistants use to plug into outside tools. You do not need to understand the protocol itself. You only need to know that it is the reason one folder can serve Claude, Cursor, and ChatGPT at the same time, rather than serving whichever one you happened to set up first.

What the setup looks like

Conceptually there are three moves, and they are the same three no matter which assistant you are pointing at the folder.

1. Choose the folder. One place on your computer where your notes live. If you already keep notes in Markdown (an Obsidian vault, for example), you are done with this step already, because those are plain files. There is more on that in connecting an existing vault.
2. Point each AI at it over MCP. Each assistant that supports MCP gets connected to the same folder. Not a copy, not a sync, the same folder.
3. Decide who gets to see what. Each connected AI is limited to the folders you allow, from a dashboard you control.

The exact clicks differ per assistant, and they change as those apps change, so this post deliberately does not invent them. The step-by-step guide for each client is provided at early access, which opens on 4 August 2026. If you want it, put your email on the waitlist at Callosium.

What changes the moment it is connected

• Your AI recalls before it answers. It pulls the relevant notes first, then answers, and cites which note each part came from.
• It files what it learns. New information goes into the right place in your folder, stamped with which AI wrote it. So when you find a claim in your notes six months later, you know whether Claude, Cursor, or ChatGPT put it there.
• It says when it does not know. If the answer is not in your notes, it tells you plainly instead of inventing something plausible.
• It works offline. Everyday recall runs on your machine using keyword search, a map of how your notes connect, and a small local model. No large language model in the core, no paid API key. The only network use is sign-in and update checks.

Who is allowed to touch what

Sharing one memory across several assistants raises an obvious question: what stops one of them from writing all over everything? Per-agent control is the answer. You allow each connected AI into specific folders and no others, and the server stamps every write with which AI made it, so the record cannot be faked. Pieces of this exist in other tools. What is unusual is having both, in something a normal person can run without a terminal.

Does it actually find the right note

Two different tests, and it matters which is which.

On the synthetic benchmark (CallosiumBench v4, 15,000 generated questions) the shipped engine scores 96.4%, taking 49ms for a typical request and 116ms for the slowest request in a hundred. That number tells you the machinery is fast and sound. It does not tell you how it handles real, messy human questions.

On the hard human bench (84 hand-written questions, graded by an independent model) a single recall call fully answers the question about 71.8% of the time. Keep the caveat attached to that figure: it sits in a roughly 72 to 79 percent band depending on who does the grading. Given a follow-up call or two, a connected AI almost always retrieves the answer that is genuinely there.

Two more numbers worth knowing. Across those 15,000 questions, including 1,450 traps about things that do not exist, the engine invented nothing at all. Hallucination rate 0%. And it reads about 3,400 tokens of context per answered question, roughly 99% less than the wasteful habit of dumping the top five matching notes in full, which runs to about 255,000 tokens.

Where it still struggles

Multi-hop questions, where half the answer lives in one note and half in another, are the genuine weak spot. On the hard human bench those score 52.9%. That is the honest number, not a rounding of it. If your question needs two separate notes stitched together, expect to ask a follow-up.

How it compares to the alternatives

• Built-in memory in ChatGPT or Claude: does not work across every AI you use, because it is one copy per vendor, and it is not plain files you can open yourself.
• Cloud memory tools (Mem0, Zep, Supermemory): usually work across assistants, but your notes are not plain files on your own machine.
• Other local tools (Basic Memory): plain files, works across assistants, but generally needs Docker or a database to get running.
• Callosium: plain files, works across every AI you use, unzip and run, no terminal.

One more comparison worth having. Against a plain keyword search agent pointed at the same folder, Callosium found the answer note 46 times out of 71 (65%) versus that agent's 10 out of 71 (14%), about 4.6 times more often.

Start with the folder, not with the tools

If you want to try this, the most useful thing you can do before early access opens is get your notes into one folder of Markdown files. That is the whole foundation, and it is worth doing whether or not you ever connect a single assistant to it, because plain text files outlive every app that reads them. There is a walkthrough in building a folder every AI can read, and if you would rather start with one assistant and expand later, the specifics are in making Claude remember your work and giving ChatGPT memory. The core is free forever and open source under Apache-2.0, so the only thing standing between you and one shared memory is the folder, and you can start that today.

Built-in memory is not memory, it is one copy per vendor, which is why you keep introducing yourself to software you have used for a year.

Frequently asked questions