I asked AI to review what it remembers about me

Because I use AI a lot, I proactively give it plenty of background.

Work history, how I’m living, projects I’m working on, personal preferences. Some of it really does make later answers fit me better.

But a few days ago a question suddenly occurred to me:

What exactly does AI remember about me?

So I asked it to review its existing memories and sort them into a few groups:

• still accurate and useful long term;
• outdated;
• not much help in the future;
• possibly too personal or sensitive;
• things I may not know were saved.

After the review, I did find that some private information that shouldn’t be kept long term had crept into its memory, along with some outdated content.

In the end, I had it delete some of it.

It reminded me:

We tend to focus on “how do I get AI to understand me better,” and rarely check the other way round —

does it know too much?

AI memory isn’t inherently a bad thing

Remembering your career background, how you like things phrased and your long-term goals can cut down on repeated explanations and make answers fit your situation better.

The question isn’t “can AI remember.”

It’s that different pieces of information differ in long-term value and sensitivity.

For example:

Suitable to keep long term:

• stable language and phrasing preferences;
• long-term career background;
• ongoing projects;
• non-sensitive usage habits.

Worth updating regularly:

• your current job;
• where you live;
• plans you’re in the middle of;
• relationships and life circumstances that have changed.

Not necessarily suitable to keep long term:

• exact addresses and real-time location;
• account and financial details;
• medical and health privacy;
• the specifics of intimate relationships;
• company, client and unreleased project materials;
• ID documents, passwords, verification codes and other credentials.

The point isn’t to build an absolute list.

It’s to ask:

Does this information really help future answers? Is that help worth the privacy cost of keeping it long term?

You can do a five-minute check too

Step one, ask the AI:

What long-term memories about me can you currently see or use? Please don’t repeat sensitive information in full; just summarize by category.

Step two, have it sort the information:

Please sort these memories into:

• recommended to keep;
• outdated, needs updating;
• not much help in the future;
• possibly too personal or sensitive;
• needs my own judgment.

Step three, decide item by item:

• keep;
• update;
• delete;
• leave for now.

Step four, don’t just trust a reply that says “forgotten.”

Different AI products may handle memory, chat history and data settings separately.

With ChatGPT, for example, turning memory off doesn’t automatically delete what’s already saved; deleting the original chat doesn’t automatically delete the saved memory formed from it. If you want something removed completely, you need to delete both the memory and the chat where you first shared it.

So after the review, you should also go into the product settings and check:

• Saved memories;
• Chat history;
• Data controls;
• Connected apps;
• permissions you no longer need.

Beyond memory, check what AI can access

AI may know more than what you’ve told it yourself.

If you’ve connected email, cloud drives, calendars, code repositories or other apps, it may also have a much wider range of information and permission to act.

While you’re at it, check:

• which apps are still connected;
• whether it can access one file or the whole account;
• whether it only has read access, or can also modify, send and delete;
• whether authorizations you no longer use have been revoked.

The more AI knows about us, the more convenient the experience may be.

But convenience doesn’t mean it should accumulate without limit by default.

A truly mature way of using it isn’t just telling AI more and more.

It also includes stopping now and then to check what’s still worth keeping, what should be updated, and what should be taken back.

I’ve put the full self-check prompt in the attachment.

It won’t ask you for the private details themselves; it first confirms what the AI can actually access, then helps you go through keeping, updating, deleting and checking permissions item by item.