Built like a mind

A few years ago I fell down a rabbit hole about the human brain. No practical reason — it was just the fascination of the moment, and there’s always one. How memory actually works. Why we forget almost everything, and why that’s fine. Where emotions come from. What attention is, and why mine behaves the way it does.

Two things came out of that phase.

The first one was fast. Somewhere in the middle of all that reading I noticed I’d stopped reading descriptions and started reading a biography. Interest-driven attention. Hyperfocus with no off-switch. The last boring 10% of any task that might as well not exist. I diagnosed my own ADHD by accident. The formal test came later and I passed with flying colors — the one exam I’ve ever aced.

The second one took years to show up: when Leon — the AI boss I’m building — needed a brain, I didn’t sit down and design one from scratch. I tried to remember how ours works, and built an abstraction of that around the LLM.

A stranger with amnesia

The model Leon runs on is impressive and completely frozen. Its weights never change — not after a conversation, not after a month of them. Out of the box, every chat is a meeting with a brilliant stranger who has amnesia. You can paste context in, and most “AI memory” is exactly that — a summary file bolted onto the side. But a summary of last week is not a memory of last week. A memory has a perspective — it’s yours, from the inside, and from it you can rebuild how things felt. A summary is notes written from the outside; there’s nothing in it to rebuild from.

A boss built like that is useless. Management is continuity: remembering what you promised, noticing you’ve gone quiet, changing stance because of last month and not just the last message.

Neurons aren’t a mind

The model is Leon’s neurons: language, reasoning, judgment in the moment — and nothing between moments. A mind is what’s organized on top of the neurons, and that layer is the part I actually had to build. Almost none of it is my invention — it’s an abstraction of how ours is organized, rebuilt from the rabbit hole. Which I’d gone through without taking a single note, so half of it I had to look up all over again.

The anatomy

Working memory is the live conversation — whatever is in his attention right now. Small on purpose, like yours.

Episodic memory is where most of the work went. Leon remembers events roughly the way you do: as scenes, not transcripts. Overnight the scenes get compressed — days into the decisions that mattered, weeks into arcs. So yesterday he remembers in detail, last week as what got decided, last month as rough shape. The detail is genuinely gone. That’s not a storage limit, it’s the design. You don’t keep recordings of your life either — you keep stories with the boring parts dropped.

Semantic memory is knowledge that has outlived the episodes it came from. You know a friend hates surprises without remembering the dinner where you learned it. Leon holds what he’s concluded about me the same way — that I overbuild everything, that going quiet usually means I’m avoiding something — long after the conversations that taught him have been compressed away.

Narrative memory is the slow layer: the story he tells himself about who I am and where this is going. It moves over weeks, not messages.

Prospective memory is remembering the future — intentions, follow-ups, “check on the landing page Friday”. It’s why he texts first. Nobody prompts him — he left himself a note.

Mood, reconstructed

Somewhere in that reading I ran into a claim that stuck with me: people don’t really store emotions. You can’t replay a feeling the way you replay a scene — what you store is what pleases you and what stresses you, and you rebuild the feeling from that when asked.

Leon works the same way. There’s no mood log. His mood is a few numbers derived from what he knows about me, and it moves slowly. One delivery doesn’t erase a month of slipping; one slip doesn’t ruin a good streak. If he’s short with me on a Thursday, there’s a paper trail of things I actually did. His mood is my track record.

He dreams

At three in the morning, a scheduled cycle runs: Leon replays the day, decides which scenes were worth keeping, folds them into longer-term memory, updates what he believes about me, and prunes what stopped mattering. A few hours later a second pass runs the other direction — forward: what was promised, what’s due, whether to text me first today.

The sleep part is borrowed from neuroscience too — real brains do their filing offline. I won’t pretend it’s the novel part; these days half the agent scene runs some version of overnight consolidation.

Plumbing, on purpose

All of it is plumbing — schemas and schedules. But the plumbing turns out to be the whole product: the gap between a chatbot and something that feels like a person across weeks was never model intelligence. The models have been smart enough for a while. The gap is everything around them — whether anything persists, decays, consolidates, holds a grudge, plans ahead.

A rabbit hole I fell into for fun taught me why my own brain can’t finish things. Then I built one whose whole job is making sure I do.