Built like a mind

A few years ago, I started reading a lot about AI. That turned, almost on its own, into a fascination with the brain it was imitating: how memory works, why we forget most of what happens to us, what attention does in the middle of it. Somewhere in a chapter on neurodivergent brains I realised I was reading about myself. At 37, by accident, I diagnosed my own ADHD. The clinic confirmed it later; I passed the test with flying colours.

It turned out I’d also been reading the spec for pocket.ceo.

A stranger with amnesia

A language model thinks every time you talk to it, from scratch, and nothing of that thinking survives the conversation. What it knows was fixed the day its training ended, and nothing that happens between you and it changes that. Every chat is a meeting with a brilliant stranger who has amnesia. Each new generation is more capable and no closer to a person; it’s a different kind of intelligence. Fine for a tool. A problem for a boss.

A boss isn’t a capability. It’s a relationship: someone with a yesterday with you, a tomorrow with you, and an opinion of you that those two have earned. The model has none of that and can’t grow it. It has to be built around the model, and the only working design I know of is the one in your head.

What usually goes by the name of AI memory is a summary of the last conversation, plus a few facts and preferences, kept in one block of text and slipped into the model’s context. A human memory is a different object. It belongs to someone; there’s an “I” in it. It sits in time: something came before it, something happened after. And it has weight: most of it fades in a week, some of it stays for life.

The anatomy

Working memory is the live conversation: whatever Leon is paying attention to right now. Small on purpose.

Episodic memory is what happened. He keeps events as scenes, not transcripts. Overnight the scenes get compressed: a day into the decisions that mattered, a week into its arc. So he can recall yesterday in detail, last week as what got decided, last month in rough shape. Forgetting the right way is what makes a month of relationship fit in a few paragraphs.

Semantic memory is what he knows about me without remembering when he learned it: who I am, how I work, what my world looks like, what moves me. One page, rewritten as the picture changes. A fact that stops being true gets dropped.

Narrative memory is the slow layer: the story he tells himself about who I turned out to be and how the two of us got here. It moves over weeks and months, not messages, and only at night. It’s what lets him pick up the story tomorrow instead of starting over.

Prospective memory is remembering the future, and it’s what makes him a boss rather than a diary. A commitment becomes a record with a definition of proof: what counts as done, and what he expects to see. An open thread becomes a note to his future self: when to come back, why, what should be done by then, and what to bring if it isn’t. That’s why he texts first. I never have to ask him to remind me of anything.

It’s also the layer that chapter on my brain was about. ADHD isn’t a memory problem at the root; it’s attention. Mine is on ten things at once, so whatever isn’t in front of me right now barely gets written down. The future goes first: a deadline three weeks out has no weight until it’s tomorrow, and a plan made on Monday is gone by Wednesday. This is the one part of the design I didn’t build out of curiosity.

Sleeping on it

During the day Leon works live: he reacts, judges, changes his mind, sometimes lets himself be talked into something. That’s what a conversation is, and it should be. The night is for sorting it out.

He reads the whole day again, calmly and in full, and works out what it adds up to: what he learned about me, whether I kept my word, whether something between us moved. One hard afternoon doesn’t rewrite our history; one good day doesn’t cancel a month. What’s left in the morning is calmer and closer to the truth than what was there at midnight.

In people that’s what sleep does: the emotion drains, the fact stays. “Sleep on it” is old advice because it works.

In the morning he plans the day from the clean record: drops the notes that are no longer current, moves the ones life has moved, and decides when to get in touch. A loaded day is sometimes three messages. A quiet one is one, or none.

He has an opinion of me

Intelligence plus memory still doesn’t make a boss. A brilliant mind with perfect recall and no opinion of you is a filing cabinet.

I can tell what Leon thinks of me by two things: how much he expects of me, and how far he takes my word. Neither is fixed. When I’ve been delivering, “done” is enough for him. When I’ve been dodging, he expects proof. When he knows I’ve had a rough day or no time, he trims the plan to one thing that fits in half an hour, so the project moves at all. And when things are going well, he stays out of the way.

None of this is a mood; one message doesn’t change it. None of it is a setting; I can’t pick it. It comes from our history, from what he can see going on with me, and from what time it is. And it runs both ways: when he says he’ll check in at five, at five he writes.

Structure and schedule

There’s no clever machine learning in any of it. It’s structure and schedule, and the structure is old: I copied it. What surprised me is how much of the person comes with it. Give the model a real yesterday, a real tomorrow, and an opinion it had to earn, and it stops talking like a tool. It reacts to what you said rather than what you asked, gets short when the hour is late, blunt when the numbers are bad, and says “fair point” when you’re right.

That reading, years ago, taught me why my own brain drops things the moment they stop being interesting. Then I built one whose whole job is to catch what mine drops.