Cross-tool setup

One memory for Claude Code, Cursor and Copilot

You fix a bug with Claude Code in the morning. In the afternoon you open Cursor, and it knows nothing about it. Each tool has its own memory, if it has one at all.

Can Claude Code, Cursor and Copilot share one memory?

Not by themselves. Claude Code keeps its notes in files such as CLAUDE.md, Cursor reads its own rules, and Copilot reads .github/copilot-instructions.md. None of them reads the others' notes.

An MCP server (a program that gives an AI tool extra functions) sits outside all three. When each tool talks to the same server, they share its memory. cachly is such a server. Its memory is called a Brain, and it stores lessons: a topic, what worked, what failed, files and commands.

How do the tools find the same Brain?

Each tool starts the cachly MCP server. The server uses the Brain named in CACHLY_BRAIN_INSTANCE_ID if that is set. Otherwise it uses the Brain saved in ~/.cachly/credentials.json.

Your key stays in that file in your home folder, not in the project. The project files hold only the server address and the Brain's ID, so you can commit them. Every tool on your machine therefore ends up at the same Brain.

How do I set up each tool?

Claude Code
Install the plugin with two commands. It adds three hooks: a briefing at the start of a session, matching lessons next to your prompt, and a lesson saved when the last message reports a fix.
Cursor
Run autopilot in the project. It writes the MCP config and the rule file .cursor/rules/cachly.mdc, and a hooks file when the project has a .cursor folder.
GitHub Copilot
Install the VS Code extension, or run autopilot. Either writes .vscode/mcp.json and .github/copilot-instructions.md.
Windsurf, Cline, Zed
autopilot covers them as well. It writes the MCP config and the rule file each one reads, such as .windsurfrules or .clinerules.

For Claude Code, type these two commands:

Claude Code
/plugin marketplace add cachly-dev/cachly-mcp
/plugin install cachly-brain@cachly

For the other editors, run this in your project folder:

terminal
npx @cachly-dev/mcp-server@latest autopilot

It asks you to sign in once in the browser. It needs no password and no credit card. Restart the editor afterwards. The full steps for every tool, with screenshots, are in the setup guide.

How do I check it works across tools?

  1. In Claude Code, ask it to remember something small, for example: "Remember that deploys go through make deploy."
  2. Open Cursor in the same project and ask: What does my cachly Brain know about this project?
  3. The answer should mention the deploy command. The agent should call smart_recall to get it. In Copilot, ask the same question.

If a tool answers without calling a cachly tool, see the troubleshooting list at the end of the setup guide.

Where do the tools differ?

The Brain is the same everywhere. How lessons reach the agent is not.

  • Claude Code gets matching lessons next to your prompt through a hook.
  • Cursor and the Copilot CLI hand lessons over with the first tool call of a turn. The details are in the same hooks in Cursor and in VS Code and Copilot CLI.
  • Some editors do not pass on the server's instructions: Cline, Continue and Zed, for example. There, the rule file from autopilot is what makes the agent check its memory.

A team can share one Brain across all members and all their tools. See shared memory for AI coding agents in a team. If you came from Cursor's removed Memories feature, start with Cursor Memories removed.

Frequently asked questions

Can Cursor read what Claude Code stored?

Yes, if both are connected to the same Brain. A lesson that Claude Code stores sits in the Brain, and Cursor finds it with the same search. Neither tool keeps a private copy.

Do I need a separate account for each tool?

No. The key lives in ~/.cachly/credentials.json in your home folder, and every tool on that machine reads it. On a second machine you sign in once with the browser.

Do all tools behave the same?

The memory is the same, the hooks are not. Claude Code gets lessons next to your prompt. Cursor and the Copilot CLI hand them over with the first tool call of a turn. Cline, Continue and Zed rely on the rule file that autopilot writes.

Why not just keep notes in AGENTS.md?

A rules file is read in full at the start and grows with every note. Lessons in a Brain are searched, so the agent gets the few that match the task. Both can live side by side.

Sources

Checked against the code and docs on .

More: MCP integration · AI memory in three layers · All docs