ChatGPT memory: what it actually remembers, and what it can't
Espen Oddvik · Founder
A founder I know was delighted the day ChatGPT remembered his product's name without being told. He mentioned it to his co-founder, who opened her own ChatGPT and asked about the product.
Blank stare. Never heard of it.
That little scene is ChatGPT memory in one act, and it's worth understanding precisely, because "the AI remembers now" is both true and misleading. Here's what it actually covers.
What memory is
ChatGPT memory does two things. It saves specific details it picks up from your conversations, like your role, your preferences, things you've told it about your work. And it can draw on your past chats when answering something new. You can see the saved memories in settings, delete them one by one, or switch the whole thing off.
For a single person, this is genuinely useful. It stops re-asking what you do for a living. It remembers you want short answers. Over months, your ChatGPT gets noticeably more shaped to you.
Credit where due: for personal use, memory solves a real annoyance well.
The three limits that matter
It's private to you. Everything the founder's ChatGPT learned lives in his account. His co-founder's account learned different things, in a different order, from different chats. There is no shared pool, so a team of ten runs ten diverging memories of the same company.
It decides for itself. You don't fully control what gets saved, what gets skipped, or what gets pulled into a given answer. Something you mentioned once in March may surface in October, and the pricing you corrected may or may not stick. Memory is a helpful reader taking notes, and sometimes the notes are odd.
Nobody reviewed it. Whatever memory holds became "true" by being typed in a chat. A guess, a hypothetical, an old number: all the same to memory. There's no step where a person confirms this is what the company actually says.
None of these are flaws for the personal case. All three become the whole problem the moment the context belongs to a company rather than a person.
The question that sorts it out
Ask yourself: should this piece of context be identical for everyone on the team?
Your preference for short answers? No. Memory's job, leave it there. The product's exact name, the pricing rules, the claims the company never makes? Yes, identical for everyone, which means it needs a home that's shared, deliberate and reviewed, none of which memory offers or claims to offer.
That shared home is what we built Company Brain to be. The company's facts live in one place, a person approves each one before it counts, and every teammate's AI pulls the same set automatically, in ChatGPT and in Claude, Copilot and Cursor beside it. Memory keeps doing what it's good at, on top of a floor that doesn't move.
The founder and his co-founder run it that way now. Both their ChatGPTs know the product.
Same spelling, even.