Three ways, compared
| Feature | By hand | A local memory server | A hosted memory |
|---|---|---|---|
| Setup | Paste into each app’s memory, CLAUDE.md, AGENTS.md or rules | Install and run a server, then add it to each agent | Paste one address into each agent and sign in |
| Shared between agents | Only what you copy, when you copy it | Yes, for agents running on that computer | Yes, every agent you connect |
| Shared between computers | Only if you sync the files yourself | Not unless you host it somewhere | Yes |
| Kept up to date by | You | Your agents | Your agents |
| Cost | Free | Free open-source servers, and your time | Saudad: a paid plan, after a free trial |
By hand
Each agent has a place for standing instructions: Claude Code reads CLAUDE.md, Codex reads AGENTS.md, and Cursor has rules and reads AGENTS.md too. Claude can also import memory you paste in from another assistant.[1][2][3][4]
It works, and it costs nothing. But every copy is out of date the moment you learn something new, and you’re the one who has to update them all.
A memory server on your computer
The MCP project’s reference memory server keeps a knowledge graph in one local file; its maintainers present it as an educational example, not a production-ready solution. Basic Memory keeps a knowledge graph in plain Markdown files and is open source.[5][6]
A local server is private and free. It is also only on the computer that runs it, and apps that connect from their maker’s cloud, as Claude’s custom connectors do, can’t reach a server on your computer.[7]
A hosted memory every agent connects to
A hosted memory sits at one address. Each agent connects to it once and signs in, and from then on what one saves, the others can read. With Saudad that’s Claude, Claude Code, Codex, Cursor, OpenClaw and other apps, and ChatGPT on its Business, Enterprise and Edu plans (read-only on Pro). You can see, edit, undo and export everything in it.
Set it up with Saudad
Sign in.
Enter your email and the 6-digit code we send. Your memory is ready in about a minute.
Copy your address.
It’s on the connect page and in your dashboard.
Connect each agent.
Follow each app’s steps: Claude, Claude Code, Codex, Cursor and OpenClaw. The ChatGPT steps cover ChatGPT on its Business, Enterprise and Edu plans (read-only on Pro).
Tell one, ask another.
Tell Claude Code you use pnpm, then open Codex in another project and ask which package manager you use.
Questions
Can ChatGPT and Claude share memory?
Not on their own: each keeps its memory inside its own app. They can share one when both connect to the same memory. Saudad works with Claude on every plan and with ChatGPT on its Business, Enterprise and Edu plans (read-only on Pro).
Do I have to stop using CLAUDE.md or AGENTS.md?
No. Keep them for instructions that belong with a project. A shared memory holds what you’d otherwise repeat in every project and app.
Is a local memory server more private?
It keeps everything on your computer, which is its strength. Saudad keeps your memory apart from everyone else’s, and doesn’t train AI on it or sell it. To search by meaning, it sends the text of your memories to OpenAI’s embeddings API.
What does Saudad cost?
€9 a month billed yearly, or €12 monthly, VAT included, after a 14-day free trial with no card.
Sources
- 1.How Claude Code remembers your project · Claude Code documentation · checked 28 October 2026
- 2.Custom instructions with AGENTS.md · OpenAI Codex documentation · checked 28 October 2026
- 3.Rules · Cursor documentation · checked 28 October 2026
- 4.Importing memory into Claude · Anthropic Help Center · checked 28 October 2026
- 5.Knowledge graph memory server · Model Context Protocol, reference servers · checked 28 October 2026
- 6.Basic Memory · Basic Machines · checked 28 October 2026
- 7.Custom connectors using remote MCP · Anthropic Help Center · checked 28 October 2026
Changes to this page
- : First published.
Saudad wrote this page and has tried to be fair to every product in it. Each fact about another product links its source, read on the date beside it.