Matt Holmes

July 2, 2026

Your AI ain’t human

It doesn’t remember you. It doesn’t remember yesterday. Whatever’s in your files is everything it knows.

One wrong label, re-read and amplified every session.

Your AI ain’t human. It doesn’t remember you, doesn’t remember yesterday, doesn’t remember the last thing it figured out. Every session it wakes up blank.

I run models from five different AI shops, on the terminal and through their APIs. Half of them can’t reliably tell me what day it is. And people trust these things to remember their business.

Two panels. Top: the model wrote 'Two days ago we announced' into a draft post. Bottom: I correct it — it was 3:31 that morning, hours not days — and it replies that I'm right, my error, fixed.
"Two days ago," in a post it was about to broadcast. It happened that morning at 3:31. No clock, no doubt.

And no, saving a ChatGPT project or a profile full of instructions doesn’t fix that. That’s not memory. It’s a sticky note you wrote once and never update.

Real memory is the files underneath. The substrate. The next session boots from them and treats every word as true.

Once you see that, you see the risk.

Label something wrong and you’re fucked. Not once. You wrote the mistake into the memory that runs every session after.

Here’s how it actually goes. A conversation gets logged in the wrong place. It’s a side project, and it lands in your company’s CRM. Now the system does exactly what it’s built to do with a company conversation. Logs a new win. Logs a new signal. Both fake. Neither has a thing to do with the company. One wrong label just bred two more wrong records, and they look every bit as legit as the real ones.

A wrong prior doesn’t sit there. It compounds.

And that’s a junk CRM row. Small. Now run the same thing through the stuff that costs you. Your pricing. Your SOPs. A strategy snapshot from three months ago that the next session reads as today’s plan and acts on.

And the worst state isn’t a gap. A blank gets flagged. The system knows something’s missing. The dangerous one is confident-but-wrong. Clean, polished, false. The system trusts it. The next session builds on it. So do you.

It muddies everything. Now you’re spending your day cleaning up a mess the system made from one bad label.

Which is why the job doesn’t go away when the task gets cheap. The AI does the work, the files hold the memory, and the part that’s still on you is keeping the state true. It’s why I build a company brain I can actually trust, then guard what gets written into it. Catch the wrong record before it ships and breeds. You’re the circuit breaker.

The task got cheap. Knowing what’s true got more valuable.