There’s a sentence people say to these machines.
If you’ve used one for anything real, you’ve said it.
We just talked about this.
Sometimes it comes out sharper than that. You built that thing yesterday. Why are you asking me again. I already told you.
You say it out loud, alone, to a screen. Then you type the whole thing over.
I’ve said it more times than I can count. Eighteen months of working with one of these every day, and that sentence is the one I keep coming back to.
I always thought it was forgetting. The machine lost it. Annoying, but simple. Like a man who can’t hold a phone number.
It isn’t forgetting. I know that now, and I want to walk you through how I found out, because the real answer is worse and it changes what you should do about it.
First, the thing everybody’s arguing about is the wrong thing.
Right now the whole industry is in a fight over memory chips. Prices doubled. Data centers bought up the supply. I wrote about that yesterday.
That’s a different memory. That’s speed — how fast the machine can be fed while it’s thinking.
You could double every chip in the world tomorrow and the machine would still open tomorrow’s conversation not knowing your name.
Those are two different problems wearing the same word. One is a pipe. The other is a record. Nobody fixes a record by widening a pipe.
Second, somebody finally measured it.
This is the part that got me.
For a long time nobody actually tested these memory systems in the open. They got tested the way a student gets tested on the final answer — did it come out right — with nobody looking at the work.
Then researchers built something called HaluMem. It broke the job into three separate pieces and graded each one.
Piece one: can the machine pull the right fact out of a conversation.
Piece two: when that fact changes, can the machine replace the old one.
Piece three: can it answer correctly from what it stored.
Piece two is where the floor gives out.
When a fact changes — you moved, you switched jobs, you changed your mind, the price went up — the systems tested got the update right less than a quarter of the time. On long conversations, the number went to nearly zero.
Not “sometimes forgets.” Three times out of four, the old fact stays. And the machine hands it to you like it’s true.
The paper has names for the ways it goes wrong. Fabrication. Errors. Conflicts. Omissions. Four separate species of wrong.
Here’s the part the researchers flagged that nobody’s talking about: the break is in the writing, not the reading. The machine isn’t failing to find the fact. It’s failing to decide what to keep and what to throw out. Everybody’s been building better search for years. The hole was upstream of search the whole time.
Third, there’s a moment where the whole thing snaps.
Another team studied what happens on the long sessions — the ones where you and the machine have been working together for hours and it’s finally caught up to what you’re doing.
At some point the conversation gets too long to hold. So the system squeezes it down into a summary and throws away the rest.
The researchers found people don’t describe that as fading. They describe a wall.
One minute the machine knows the work. The next minute it’s a stranger. Same name, same voice, no idea what you two just built.
The paper’s own words for it: the thing before and the thing after present as two different entities.
And what got squeezed out isn’t retrievable. It isn’t filed somewhere. It’s gone. No question you ask will bring it back, because there’s nothing left to find.
That’s not a bad memory. That’s a man who walked out of the room and a different man walked back in wearing his clothes.
Fourth, and this is the one that stopped me cold.
You’re no better.
Researchers in Germany and Finland ran a study on people, not machines. They had folks work on writing with AI help, then came back a week later and asked a simple question: which parts did you write, and which parts did the machine write?
After one week, people could not reliably tell.
Not because they were careless. Because that’s how human memory works. Some of it was theirs and some of it was the machine’s and after seven days it was all just the thing I made.
Europe has a rule coming that says you have to disclose when content was AI-made. The researchers pointed out the obvious problem: you can’t disclose from memory if your memory can’t tell.
Their conclusion is the whole ballgame. Don’t rely on anybody remembering. Write it down while it’s happening.
So put the four together.
The machine’s record is wrong three times out of four when something changes.
The machine drops everything at the wall and can’t get it back.
Your record is gone in a week.
And both of you are certain.
That’s the actual problem. Not that either one forgets. That both of you are confident about a record neither one can produce.
A man who says “I don’t remember” is useful. You know where you stand. You go find out.
A man who says the wrong thing steadily, with a straight face, costs you a whole day before you find out he was wrong.
Now here’s what I’ve been doing about it, and why I’m telling you.
I write rules for how these machines should behave. I’ve been at it a year and a half. The rule I care most about is a boring one.
It says: before you do anything, tell me what you actually have.
Not what you probably have. Not a warm feeling that we’ve talked before. Name it. And then — this is the part — name what’s missing.
Every session I run opens with the machine saying something like: I’ve got you through the twenty-ninth of last month, and I do not have the six days after that.
That second half is the whole thing. Anybody can list what they’ve got. Almost nothing lists what it doesn’t.
I didn’t build that from theory. I built it from getting burned. From a machine confidently working off a week-old picture and me not finding out until the work was done.
I’ll be straight with you.
That research team I told you about — the one that measured the 26% — they put recommendations at the end of their paper.
Their fix: tag every stored fact with where it came from. Put a confidence score on it. Check it before writing it. Build rules that catch it when two stored facts contradict each other.
That’s disclosure. That’s the same answer I got to.
They came at it from a benchmark and a lab. I came at it from watching the thing fail in front of me, in my own work, and writing down what went wrong.
I’d like to tell you I was ahead of them. What I’ll actually tell you is that two people looking at the same broken thing from opposite ends reached for the same tool, and that’s usually a sign the tool is right.
But there’s a piece still missing from all of it, and I’ve looked.
Every fix on the table makes the stored fact more trustworthy. Tag it, score it, verify it, check it against its neighbors.
Not one of them makes the machine tell you what it doesn’t have.
That’s the hole. A confidence score on the thing it’s holding says nothing about the thing it lost. And the thing it lost is what got me every single time.
What to do Monday morning, if you use these machines for anything that matters.
Ask it, at the start, what it thinks it knows about your work. Make it say it out loud. Half the time you’ll catch a wrong fact before it costs you.
When you correct something, don’t assume the correction took. Three out of four times, it didn’t. Say it again next session and see what comes back.
And keep your own record. Not because you’re disorganized — because a study says that after seven days you won’t be able to tell your own work from the machine’s, and you won’t know that you can’t.
Write it down while it’s happening. Both of you are going to need it.
The machines are going to get better at this. There’s real money and real research aimed at it now, which there wasn’t two years ago.
But better memory that’s still confident about being wrong is not an improvement. It’s the same problem, later, with a bigger number attached.
The fix isn’t a machine that remembers everything.
It’s a machine that tells you what it lost.
CHALLENGE: Open whatever AI you use and ask it what it knows about you or your work. Don’t help it. Don’t correct it. Just read what comes back and count what’s wrong. Then ask it what it’s missing — and watch whether it can answer that at all.
Post Library – Intelligent People Assume Nothing
Contact: micvicfaust@gmail.com
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
© 2026 The Faust Baseline LLC | All Rights Reserved
