If you ask an AI assistant "What do you know about me?" or "What do you remember about me?", its answer comes from a short list of places. Those are facts it saved to memory, the instructions or profile text you wrote, past chats it is allowed to reference, whatever is in the current conversation, and any apps or files you connected. Outside that list it knows nothing about you personally, unless you are public enough to appear in its training data or in a web search. You can check everything on the list, and you can delete most of it.
Where the answer comes from
The first source is saved memory. Assistants such as ChatGPT and Claude can keep facts between conversations, like your job, a project you keep coming back to, or how you like answers formatted. Depending on the product and your settings, some of these get saved automatically and some only when you ask. This is usually most of what you hear back when you ask the question directly.
The second is instructions you wrote. Custom instructions or profile preferences are text you typed into settings at some point. The assistant reads them at the start of every chat, so they show up in its answer even if you have forgotten writing them.
Then there are past conversations. Some assistants can search or reference earlier chats when a setting allows it. That works separately from saved memory, and whether an assistant remembers your other conversations depends on which of the two you have switched on.
The current chat counts too. Anything you pasted, uploaded or said earlier in this thread is in view for as long as the thread lasts.
Connected tools are another source. If you added a connector for email, documents or a calendar, the assistant can look things up there when a question calls for it. It doesn't read all of it in advance. It searches when asked, and only within what you authorized.
Last is public information. If you have a public profile, the model may know something about you from training data, or it may find something through web search if that is on. This source is the least reliable, and it is the one most likely to produce confident wrong details.
How to check it properly
Asking the question gives you a quick first look. The answer is a summary the assistant writes on the spot, though, and summaries leave things out. The settings page is the better record. Look for a memory or personalization section that lists saved items and lets you remove them one at a time, and look for the switch that controls whether past chats are referenced. Check your list of connected apps as well, because a connector you added months ago keeps working until you remove it.
If you want a conversation that leaves nothing behind, most assistants have a temporary or incognito mode that skips memory. Deleting a chat removes it from your history. A fact that was saved to memory from that chat may stay, though, until you delete it on its own.
What it does not know
What people tend to find most surprising is how little the assistant knows about their own work. It can know your job title and how you like your emails written and still have no idea what is in the recordings and project folders you open every day. Nothing in memory describes them, and it can't open a file you never shared.
Video is the clearest case, because nothing about what is said or shown in a video can be searched until something has looked inside it. With Vivu, you upload videos to a project where they are indexed once in the cloud, and then Claude, ChatGPT or another assistant that supports remote MCP connectors can search that project when you ask about a moment. You get back time segments, sometimes from several different videos, each with a short reason it matched, and a few will be only loosely related, so you still decide which ones count.
The same rule holds for any connected source: the assistant knows what it can search, at the moment it searches.
When you don't need to dig into this
Say you use an assistant for one-off questions, keep memory off and have connected nothing. Then "what do you know about me" really does come down to what is in the current chat, and there is nothing to audit. Checking starts to matter once memory is on, once work accounts are connected, or once you share an account or a device with someone else.
How to tell which case you are in
Ask the assistant the question, then open settings and compare the two. If they match and nothing is connected, you are in the simple case and can stop there. If settings show more than the answer did, go by settings and prune from there. And if what you were hoping for was an assistant that knows your work, memory won't give you that: it only knows your files and footage through sources you connect and it searches.
FAQ
Does an AI assistant still remember things after I delete the chat?
It can. Deleting a chat removes it from your history, but a fact that was saved to memory during that chat is stored separately and may remain.
After deleting anything sensitive, open the assistant's memory or personalization settings and remove the matching saved item there too.
Can ChatGPT or Claude see my email or files if I never connected them?
No. An assistant can only reach email, documents or other files through an upload or a connector you set up. Even then it searches when a question needs it rather than reading everything in advance. Removing the connector ends that access.
Why doesn't the assistant's answer match my memory settings?
The answer is a summary generated at the moment you ask. It can skip saved items, and it can mix in details from your custom instructions or the current chat. The list in your memory settings is the actual record, so use it when you want to know exactly what is stored.
Does the assistant know about me from the internet?
It only does if information about you is public, either because it appeared in the model's training data or because the assistant ran a web search. That kind of knowledge is often out of date or attached to the wrong person with the same name. Treat it as unverified and correct it in the chat if it matters.