ChatGPT Business gained admin controls for shared project memory so companies can limit what sticks across chats.
OpenAI gave ChatGPT Business admins a proper dial for shared project memory. Personal memory and project memory are no longer a single blunt switch for many workspaces. That matters once a company mixes private drafting with client folders that should not remember last quarter’s rumors.
What admins can do now
Owners can allow, restrict, or clear project-level memory. Personal memory can remain available for individuals while project spaces stay stricter. Clearing a project’s memory is a deliberate action, which is what you want when a folder held temporary credentials or a deal that died.
OpenAI’s framing is separation. A project is not a diary and not a wiki. If staff treat project memory as the system of record, you will eventually need a real document store with ownership. The admin controls reduce accidental leakage between modes. They do not replace records management.
Rollout advice
Decide which projects may remember facts at all. Client delivery projects should start restrictive. Internal brainstorming can be looser if policy allows. Tell staff when a clear happens so nobody thinks the model “forgot how to write” when it actually forgot a project’s cached brief.
Also ban secrets from prompts as a habit, memory on or off. Memory features make sloppy secret handling worse, not better. OpenAI shipped the control. Your acceptable-use note should ship the same week.
Example policy that works
Default new client projects to memory off. Allow memory on internal labs only after a named owner accepts residual risk. Quarterly, clear projects that have been inactive for ninety days. Train staff that “remember this API key” is a firing offense whether memory is on or off. OpenAI’s toggles make the policy enforceable. They do not invent the policy. Write it in one page, link it from the Business workspace description, and keep the version date visible.