When we talk about what an AI knows about you, we conflate two mechanisms that have nothing in common. Memory: information stored on your account, which the system re-reads at each conversation. Training: your exchanges used to improve the model itself. The former concerns you alone and can be controlled. The latter concerns everyone and is far harder to control.
This is probably the most frequent question from users, and the one that gets the vaguest answers. Let's clarify.
Memory: what is stored on your account
Assistants now retain information from one conversation to the next: your job, your format preferences, ongoing projects, facts you have mentioned.
Technically, it is not the model that remembers. The model is fixed, as we explained in our article on LLMs. What actually happens is simpler: this information is stored in a database and re-injected into the context at the start of each conversation, like a summary the system re-reads before talking to you.
Practical consequence: this memory is viewable and editable. Interfaces generally let you see what is retained, delete items, or disable the feature. It is the most useful setting and the least used.
Training: what is far harder to undo
This is the other mechanism, and it follows different rules depending on the type of account.
On consumer plans, conversations may be used to improve models, unless you disable that option where it exists. On business plans, providers contractually commit not to do so, which is one of the main reasons to pay for those tiers.
The difference is fundamental. Information deleted from your memory disappears. Information that has contributed to training a model is not easily removed: it would require retraining, which is neither free nor immediate.
We systematically underestimate what our exchanges contain. A technical question reveals your job and your level. A request for rewording reveals an ongoing conflict. A series of medical questions reveals a concern. Explicit content is only part of the information: the structure of your requests often says more than their content. This has been true of search engines for twenty years, and AI amplifies the phenomenon because we talk to it more freely.
The settings that really matter
Memory. Look at what is stored. You will often be surprised by what has been retained, and sometimes by what has been misunderstood.
Use for training. This is the most important setting. It is sometimes enabled by default and not always easy to find.
History. Retention duration varies, and deleting it does not necessarily mean immediate erasure from backups.
Connectors. If you have linked your email or documents via a protocol like MCP, the scope of what is accessible goes well beyond what you type.
Your rights, in Europe
The European framework applies fully to these services. You have a right of access to data concerning you, a right to rectification, a right to erasure, and a right to object to certain processing.
In practice, exercising the right to erasure over data used for training raises real technical difficulties that regulators and companies have yet to finish arbitrating.
What to take away
A simple rule beats fine-grained settings management: do not entrust a consumer assistant with anything you would not put in a professional email. This is not paranoia; it is the same principle as with any online service.
And for anything truly sensitive, the only solid guarantee remains the one we described in our guide to local AI: a model running on your own machine lets no data out, because there is nothing to let out.