What can an AI connected to my memory do?
Connecting ChatGPT or Claude to your company memory raises the right question immediately: what is that AI allowed to do with my data?
What it reads
Whatever the token's scope allows, inside the memory it is tied to. A read token gives access to the project context, decisions, learnings, recurrences and contacts, with their evidence.
In practice, the AI stops asking who you are and what you already settled. That is the whole point: it works from your history instead of reinventing it every conversation.
What it can propose to write
With a read and write token, the AI can propose adding a decision, a learning, a contact, an event.
That does not mean it adds them. A write coming from an AI follows exactly the same path as an import: it lands in Sources as a proposal, and only enters the memory once you approve it. An AI telling you it has "saved" something is describing its request, not the result.
What it cannot do
- Write straight into the memory without validation.
- Reach outside the memory its token is tied to.
- See a memory you cannot see yourself.
- Get around the rate ceilings, 30 calls per minute and 400 per hour.
How this differs from ChatGPT's own memory
An assistant's memory is tied to your account with that provider: it retains things about you, inside that tool, and crosses neither people nor other tools. The day you switch assistants, it stays behind.
A company memory is the opposite: it is yours, shareable with your team, and plugs into whichever assistant you use. That is also what lets you change assistant without losing anything.
Didn't work?
- The AI says it saved something you cannot find. That is the expected behaviour: the proposal is waiting in Sources. See Approving or rejecting a proposal.
- You want to limit what it does. Generate a read-only token and replace the old one in your client.
- You have lost track of which token is wired where. Name your tokens when creating them, one per client.
FAQ
Can I give read-only access?
Yes, that is the scope to pick when creating the token, and the one to prefer while you are finding your feet.
Does the AI provider keep what it reads?
See Is my data used to train an AI model. The short answer is no.
Can several people connect the same memory?
Yes, each with their own token. Since rate ceilings are per token, this also stops one person blocking everyone else.