Can you read what your AI remembers about your company?

Questions about AI memory at work tend to circle the same few points. Here they are, with straight answers.

Don't the big AI tools remember things now?

They do. Through 2025 the major assistants all shipped memory. ChatGPT started referencing past chats in April 2025. Anthropic extended Claude's memory to Team and Enterprise accounts in September 2025, with memory kept separate per project. Microsoft added memory to Copilot the same year.

For one person working alone, that is a real improvement. You stop retyping your role and your preferences every morning.

So what is the problem?

Two problems. The first shows up in the vendors' own community forums.

The first is that you cannot always tell what is in there. Users on OpenAI's forum have reported memory filling up with no warning, so new entries silently failed to save. Others reported the assistant saving memories they never asked for. In November 2025 a wave of users reported that their saved memories had been wiped.

The second is scope. These features remember one user, so a correction made in one account stays there. That gap matters too, but this post is about the first one.

Why does that matter more for a business than for a person?

Because a business has to answer for what its AI says. If a client receives a wrong price, someone will ask where it came from. "The assistant must have remembered something" is not an answer you can give a client, an auditor or your own team.

A firm needs to be able to do four things with its AI's memory: read it, change it, see who changed it and when, and roll a bad change back. Most built-in memory features offer the first two, partly. Few offer the last two at all.

What does an auditable memory look like?

Ours is not clever. It is a set of plain text files holding our rules, the facts we have checked, and the corrections we have made, each with the date and the reason. The files sit in a version-controlled repository, the same kind of system software teams use for code. Every change is recorded: what changed, when, and by whom. Anything can be compared with last month's version or put back.

Our AI sessions read those files before they start work. Anyone on the team can open them in a text editor. When the AI says something odd, the first question is "what does the memory say," and the answer is on screen in seconds.

Is there anything this does not solve?

Plenty. Readable memory can still be wrong. If someone writes down a bad rule, every session follows the bad rule, consistently and with confidence. Somebody has to own the memory and keep it current, and that takes real hours each month.

It also does not replace the vendor features. Personal memory for personal preferences is fine. The company's rules and facts are the part that should not live there.

And a memory everyone can read is a memory everyone can read. Keep passwords, keys and anything you would not show the whole team out of it, and split it by client or company where that line matters.

Who should own it?

One person, named, with time set aside for it. In a small firm that is often the operations lead, because they already know which facts change and when. Their job is to review changes, retire facts that are no longer true, and make sure a correction made on Tuesday is in the memory by Wednesday. They do not have to write every entry. Without an owner, a readable memory slowly fills with entries nobody trusts and nobody dares delete.

Where do we start?

Open whichever AI tool your team uses most and ask it, plainly: "What do you remember about our company?" Write down the answer, including anything it says that you did not know it had kept. Then ask a colleague to do the same in their account. Put the two answers side by side. Wherever they disagree, or wherever one of them is wrong, you have found something that belongs in a shared, written memory instead.

Book 20 minutes at pgco.world/book.