Shared CLAUDE.md, AGENTS.md and AI memory for game teams
For studios where developers work with Claude Code, Codex or Cursor. Gameheim keeps the instructions and the memory of those assistants in one place, for the studio, each project and each person, puts a person in charge of every shared change and writes the same version on every machine. This guide explains how it works and how to set it up.
Why AI instructions drift apart on a team
Every developer who uses an AI coding assistant ends up with a file of instructions, such as CLAUDE.md or AGENTS.md, and a memory the assistant keeps adding to. It works well for one person. On a team it drifts:
- Different copies. Each machine has its own version of the instructions, and nobody knows which one is right.
- Memory nobody reads. The assistant writes down what it learned, and the notes pile up. A wrong line is repeated in every session that follows.
- No author, no history. When an entry turns out to be wrong, there is no way to tell who added it, when, why, or what it said before.
- Knowledge that stays on one PC. What one developer's assistant learned about the build never reaches the rest of the team.
Committing the files to the repository fixes the copies, but not the review: a change to the team's instructions goes in with the code, and the memory an assistant writes for itself stays on the machine.
How Gameheim keeps one reviewed version
The page AI Context in Gameheim holds three levels, the same ones the tools already have. Each level has a document of instructions and a memory made of short entries, each with a title and a text in Markdown.
| Level | Reaches | Changes |
|---|---|---|
| Studio | The assistant of every member, in every project. | Through a request that a person approves. |
| Project | The assistants of the project's members. An entry can be meant for some teams of the project only. | Through a request that a person approves. |
| Personal | Your assistants only. Nobody else in your studio can read it. | Saved right away, with its history. |
- Author and history. Every version keeps who wrote it, who approved it, when and why. Restore this version brings back an old text by adding a new version: nothing is lost.
- Checked before it is saved. On the studio and project levels Gameheim refuses texts with the access keys, private keys and hidden characters it recognizes, so they do not reach every assistant of the team.
- Conflicts. If an entry changed after a proposal was made, the reviewer sees the text as it was, as it is now and as proposed, and writes the final one.
The files on disk and version control
Gameheim never changes the CLAUDE.md, AGENTS.md or Cursor rules your team keeps in the repository: it writes its own files next to them, and the tools read both.
- Git. The agent adds its files to the local
.git/info/excludeof the clone, never to.gitignore. - Perforce. The agent writes only where the files are ignored: by a block in the project's
.p4ignore, or, with your consent, by an ignore file of the agent added toP4IGNOREon that computer. Otherwise it skips those workspaces and says why. - A file that ends up tracked is not written again until it is removed from version control, and the page shows it.
- Edited by hand. If someone edits a generated file, the agent asks whether to send the changes to Gameheim (Send) or Discard them: the edits become a normal request, and the official version comes back either way.
If something cannot be removed right away, for example because Perforce is unreachable when you sign out, the agent keeps trying every few minutes. Its Left on this PC panel shows what is still there and how to remove it by hand.
The memory commands
The commands run only when you launch them in your assistant, and they work through the Gameheim MCP server. Gameheim and the agent never read your assistant's own memory: the assistant does, with you.
| Command | What it does |
|---|---|
gameheim-memory-init | The first import: goes through the assistant's memory with you and saves or proposes each entry you confirm. |
gameheim-memory-update | The same, only for what changed since the last time. Your assistant suggests it after a week, or when it saves a new rule in its own memory. |
gameheim-memory-leave | Copies your personal entries back to your assistant. Claude Code writes each one in your own rules or in its memory of the project, as you choose. Codex shows you the lines and adds them at the end of ~/.codex/AGENTS.md only after you say yes. Cursor keeps rules in its settings, so it shows you the entries to paste there. It adds without removing anything, and deletes nothing in Gameheim. |
Your personal memory
The Personal tab holds what your assistants remember about how you work. They save it without waiting for a review, and only you can see it: nobody else in your studio, the LEAD included. It has the same history and restore as the shared levels.
- Not reviewed for 90 days. Entries nobody has changed or confirmed for 90 days are marked, so old notes do not linger. Still valid confirms one; Archive stops it reaching your assistants and keeps it in its history.
- Share what helps the team. Propose to share turns a personal entry into a request to the studio or to a project. Tick Archive my personal entry when the proposal is approved and your copy steps aside once the shared one exists.
- A warning, not a block. If a personal text looks like it contains a secret, Gameheim says so and asks before saving it.
- Start over. Empty my personal memory deletes every entry, your instructions and their history for good, after a confirmation.
- A digest of what changed. Once a week by default, a notification tells you how many entries your assistants added, changed or archived, only when something changed. Frequency, email and phone notifications are in Settings, under Notifications. It carries numbers only, never the text of your memory.
Frequently asked questions
Does Gameheim change the CLAUDE.md or AGENTS.md in my repository?
No. It writes its own files next to them and never touches the ones your team versions. For Codex it copies your AGENTS.md into the file it writes, so both reach the assistant.
Will these files end up in Git or Perforce?
No. The agent excludes them from Git and writes them in Perforce workspaces only where they are ignored, and it stops writing a file if someone adds it to version control anyway.
Can my LEAD read my personal memory?
No. Only you can see it. Nobody else in your studio can read it, not even the admin.
What happens to my personal memory if I leave the studio?
It belongs to the studio's workspace, like the rest of your work there. If your account is deactivated it is kept for 12 months and comes back if you return; after that, or when the account is deleted, it is deleted. To keep something for the team, propose it to the project.
Does Gameheim read my assistant's own memory?
No. Your assistant reads it, only when you run one of the memory commands, and every entry it brings into Gameheim waits for your confirmation.
Which assistants does it work with?
Claude Code, Codex and Cursor.
Does the agent start writing files on its own?
No. Each person switches it on in Gameheim, and the agent asks once on each computer.
Can I use it without the agent?
Yes: an assistant connected to the Gameheim MCP server can read the current context and propose changes. The agent is what keeps the files on disk up to date.