What Does Your AI Think It Knows About You?

"What does it think it knows about me?"

If you've had memory switched on for a while, there's an answer, and it's longer than you'd guess. Memory is the feature that lets the AI remember things about you between chats. Your AI has built a picture of you from your job, your projects and the things you keep asking about at 11pm. The AI uses that picture to shape every reply.

Most people have never read it, and I think that's backwards. A description of you that you've never seen is steering your answers, and it was written by something that guesses for a living.

Think of it as a colleague's notes about you, which you're allowed to read.

Read the notes

Both Claude and ChatGPT let you see what they've remembered. In Claude, look under Settings, then Memory, and ChatGPT has a similar page under Personalisation. Read the whole thing before you change a word.

Each line is one of three kinds. Some things you told the AI directly, and those are usually right. Others the AI worked out for itself, and that kind of guess can be wrong but stick anyway.

The third kind was true once, like the job you left, the project you finished, or the diet you gave up in March. Notice which line is which kind, and leave the editing for later.

On some plans, searching across your past chats is a paid feature, so on a free plan the AI works from the memory page alone. That gives the AI less to go on, but the page is still worth reading.

Ask where the notes came from

Now open a fresh chat and ask the AI to describe you. The most important word in this prompt is "evidence".

From what you remember about me and our past conversations, describe who you think I am. Cover my work, what I care about, how I like to be spoken to, and what I'm usually trying to do when I come to you. For each point, say whether I told you directly or you worked it out, and give the evidence. Don't flatter me and don't fill gaps with guesses.

Read the answer slowly. Some of it will be spot on, and some will be a fair guess from thin evidence.

Watch for a third kind. This is a version of you from one bad week, when you asked a lot about one thing and the AI decided that was you. The AI's picture is built from the questions you bring it, and that's a strange way to get to know a person.

The AI has never seen you with your friends, at your best, or on a Sunday, but it will still sound sure of itself. That's why you asked for the evidence.

Correct the record

The notes are about you, so they're yours to edit. Go back to the memory page and do two things.

• Delete what's out of date, what's wrong, and what you'd rather the AI didn't keep. Wrong guesses go even when they flatter you, because being called "highly organised" doesn't help if every plan the AI writes then assumes you'll stick to it. Health, money and family deserve a second look. You can talk about them without them being stored.
• Add what's missing, in your own words. Say how you like answers, for example short and with no lists, and say what you're working on now. Say what the AI must never do, such as send things, buy things or pretend to be you.

Then say it in a chat as well. For example, "Update what you know. I've left that job, I'm now doing X, and I prefer Y." Come back to the page every three months, because the person it describes should keep changing, just as you do.

Ask one hard question, then argue

With the notes clean, ask the AI one thing worth asking.

Look at the questions I've brought you over the last few months. What have I asked about more than once without ever acting on? Point to the actual conversations, and tell me what you think is in the way.

Or, if you'd rather look forward, try this.

What's one thing I keep saying I'll do that the pattern of our chats suggests I won't, unless something changes? Be specific and show your evidence.

Then argue with it. If the AI names something you think is wrong, say so, and ask for the conversations it's drawing on.

Whatever the AI says is a colleague's opinion, never a final judgement. If anything rings true, take it to someone who knows you, and ask them. And a chatbot isn't a therapist, so if this turns up something heavy, take it to a person.

Sometimes the AI backs down, which tells you it was guessing. Sometimes it shows you three chats from the last two months, and you go quiet for a minute.