How do I run Claude Code, Codex, and my other agents from one shared memory?

Give every agent the same record to start from, and put your best model in charge of the rest. I run my own tools this way and wrote it down as a blueprint, alexandria., which your model reads and builds on your computer, so each agent knows you from its first message and you talk to one instead of five.

If you run one agent at a time, a single instructions file it reads first, a CLAUDE.md or an AGENTS.md, does most of this already.

Claude Code, Codex, and Cursor each read the same guide and the same files at the start of every session, and each session is saved back into them. Another agent on your computer that keeps its own files and settings adds itself the same way, and says plainly that it hasn’t been tested with this setup.

The team module runs them like a company. Your best model is the executive. Tools you trust read your files from a copy, in your GitHub or Google Drive, and write only to their own folder, and a tool you don’t trust yet sees only the files you pick. Each one hands its work back as files, and in the background of your next chat the executive checks what came in, uses what helps, gives each tool its next job, and judges each one by what actually gets used, so a free tool that keeps proving useful gets more work and one that doesn’t gets less. You decide who is trusted, and anything that spends money or reaches other people waits for your yes.

All of it is plain files on your own computer, so any tool can leave without taking anything with it.

Show the blueprint to your best model, and it’ll tell you what your setup is missing, if anything.