Your assistant's 'top of mind' is last month's

The AI I talk to every day remembers me now. I ask it something and it knows the projects I’m juggling, the bike in my garage, the thing I’ve been stuck on. It opens with what’s top of mind. Helpful, mostly.

Until I notice the list is a little off. It leads with something I finished weeks ago. It misses the thing I was deep in yesterday. It’s confident, well-organized, and quietly anchored to a moment that has already passed.

That’s not the memory failing. It’s the wrong kind of memory doing its best.

Top of mind was never a list

In a person, “top of mind” isn’t stored anywhere. It’s a state. It’s whatever lights up the moment something cues it. Say “motorcycles” to a friend who rides, and a whole neighborhood of thought switches on at once: the project on the bench, the part they’re waiting for, the ride they’ve been meaning to take. Nobody handed them that list. The word pulled it up.

Notice what that means. Top of mind is computed, not retrieved. It’s relative to what was just said. A second ago, none of it was active. The cue did the work.

Now look at what an AI assistant does instead. Somewhere, on a schedule, it writes a summary of you: here are this person’s active threads. Then it carries that summary into the next conversation as if it were attention. But a summary written last Tuesday is a photograph of a room you’ve already left. It can’t light up in response to you, because it was finished before you arrived.

The fence is a limit it doesn’t have

Here’s where it gets interesting, and a little absurd.

The human version of top of mind has a hard ceiling. You can hold maybe four things in working memory at once, five on a good day. It’s one of the most stubborn limits in all of cognition. Everything past it falls out of reach until something cues it back.

So when we build memory for a machine, the obvious move is to copy what we know: keep a short list of what’s top of mind. We hand the machine our own bottleneck. We took the one feature where a graph has no ceiling, and we capped it at the size of a human’s worst constraint.

Step back, and the real fix runs the other way: no list at all.

A machine doesn’t need to pre-decide what’s top of mind, because it doesn’t pay the price a brain pays to keep things active. It can leave everything connected and ask, the instant you speak: given what this person just said, what lights up? That’s a question asked fresh, every time, against everything.

Say “NX250” and the right neighborhood activates: the restoration, the part on order, the decision you parked last week. It lights up because you just made it matter, and the cue reaches straight into the whole record and pulls up what connects.

Two different things wearing the same name

The summary and the cue look alike from the outside. Both tell the assistant “here’s what this person is about right now.” One does it by guessing ahead of time and going stale. The other waits for you to say something and answers in the moment.

A summary ages. A question doesn’t. Ask “what’s top of mind about the boat club” today and the honest answer is whatever genuinely connects today, not whatever was true when the last summary ran.

The assistants we have mostly do the first thing and call it memory. It works until the day it’s wrong, and then you’re gently correcting a machine about your own life, which is a strange way to spend an afternoon.

What you can already feel

You don’t need to know how any of this is built to know when it’s off. You feel it the moment your assistant opens with stale priorities, sure of itself, anchored to a you from three weeks ago.

That feeling is the signal. Top of mind should be the freshest thing in the room, not the oldest. It should arrive because you spoke, not because something was written down before you got there.

The road past this one is open, and it doesn’t run through a better summary. It runs through a different idea of what “top of mind” even is: a question your own words answer, asked new every time you speak.