Claude Code
Connect Claude Code to Memory Bridge and it reads from the memory of your work. What you save in Claude Code is there in your other AI tools.
Guide
Memory for coding agents: Claude Code, Cursor and Codex
A coding agent starts each session not knowing what was decided in the last. Instruction files solve half the problem. This is the other half.
Last reviewed:
The short answer
Coding tools have two layers of memory: instruction files in the repository that describe conventions, and automatic memory the tool builds for itself. Both belong to one project and one tool. For architecture decisions and their reasons, and for what gets decided outside the editor, you need a shared memory every tool reads.
What each tool remembers on its own
| Tool | Instruction file | What it builds itself |
|---|---|---|
| Claude Code | CLAUDE.md files in the project and your account | Automatic per-project memory on your machine |
| Cursor | Rules files inside the project | Memories from your chats with it |
| Codex | An AGENTS.md file in the repository | Relies on the file and the task context |
| GitHub Copilot | An instructions file in the repository | Memory scoped to the repository |
| Windsurf | Rules you write for it | Memories it generates for the workspace |
We reviewed this description in October 2026. These tools change monthly, so check their official docs.
What instruction files do not carry
The instruction file is the right place for commands and conventions: how to run tests, how to name files. It should stay short, because it is read in every session.
Why that library was rejected, what was tried and failed, what the client was promised: none of that belongs there. The file grows, then goes stale, then the team stops updating it.
What a shared memory adds
- Decisions with a date and a reason, read when needed rather than every session.
- Context that crosses tools: what you discussed in ChatGPT reaches Claude Code.
- Context that crosses repositories: one convention holds across all your projects.
- Earlier versions: you know who changed a decision and when.
- Memory not tied to a machine: what you saved at your desk is on your other computer.
In an example
You
Agreed with the client: we never store card numbers ourselves. All payment stays with the payment provider.
ChatGPTSaved.
Saved to memory · Client requirementsOne memory
You
Add a "save card for next time" feature.
With Memory BridgeImplemented with a token issued by the payment provider. I did not store the card number in our database, in line with your agreement with the client.
From memory · Client requirementsIllustrative example
How to start
Keep the instruction file short, for commands and conventions. Move decisions and their reasons into memory. Then connect your tools to it: Claude Code, Cursor and Codex.
Common questions
Does memory replace CLAUDE.md or AGENTS.md?
No. The file is for conventions read every session; memory is for what gets looked up when needed. They work together.
Does memory increase token usage?
Only what the request needs is retrieved, not everything saved. That is lighter than a long instruction file read in full every session.
Does it work for a team?
Memory Bridge is built for personal use today. Team memory is not available yet.
Keep reading
Cursor
Rules say how you write code. Connect Cursor to Memory Bridge and stop re-explaining your work.
Engineering decisions
Teams change; decisions should remain. Save each technical decision with its reason and review date, so your AI tools find it when the question returns.
One memory. Every tool you use.
The trial is invite-only for now. Request access and we will write to you when your seat opens.