What each one keeps
From each company's own documentation:
| App | Its own memory | Shared with the others? |
|---|---|---|
| Claude Code | CLAUDE.md files you write, plus auto memory it writes, in a folder for each project on your computer (Anthropic) | No |
| Cursor | Project rules in .cursor/rules, user rules for all your projects, and AGENTS.md (Cursor) | No |
| Codex | AGENTS.md, plus memories it can write to ~/.codex/memories/, off until you turn them on (OpenAI) | No, not even with ChatGPT |
So a preference you teach Claude Code on Monday is new to Cursor on Tuesday and to Codex on Wednesday. Developers who switch between them end up repeating the same things three times, or keeping three files in step by hand.
Project rules: one AGENTS.md for all three
For rules about the project itself, such as how to build it, how to run the tests and the conventions to follow, there is now a simple answer: AGENTS.md. Cursor reads it as one of its kinds of rules. OpenAI suggests keeping the guidance Codex must always follow in it. And Claude Code reads it as the project's instructions when the project has no CLAUDE.md (from version 2.1.277).
One file in the repository, checked in with the code, and all three tools start from the same rules. If you already have a CLAUDE.md, Anthropic's documentation shows how to point it at AGENTS.md so the two never disagree.
What they know about you: one shared memory
AGENTS.md is the wrong place for facts about you. They aren't part of the project, they change, and they belong in every project, not one. That is where a shared memory comes in: a memory outside all three tools that each connects to through MCP, the standard way AI apps plug into tools and data.
Zaaheen is one such memory, for Windows and Mac, kept encrypted on your own computer. Each tool is connected once: Claude Code with one command, Cursor with an install link, and Codex with a few lines in its settings file (Claude Code, Cursor, Codex). After that, what any of them saves about you, the others can recall.
Our test
We tried this with Cursor reading a shared memory that held about 15 facts about a person, saved earlier by another app. The questions were the kind you would really ask, and two were designed to catch it out: one about something the memory didn't hold, and one trick question.
| Option | Questions |
|---|---|
| Recalled the fact | 3 of 5 |
| Said it didn't know | 1 of 5 |
| Avoided the trick | 1 of 5 |
Source: Zaaheen's own test, 5 June 2026, a development version of Zaaheen on Windows. Recalled: the instrument the person plays, their pet and what they collect. Didn't know, correctly: their blood type, which the memory didn't hold. The trick: asked about their cat, it said they have a dog.
All five answers were right, including the two where the right answer was "no". For a memory, that second part matters as much as the first: an app that invents an answer is worse than one that says it doesn't know.