The road it forgets

A few weeks ago I settled something with an AI. We’d worked through a design decision together, turned it over, found the reason it had to go one way, and moved on. Ten minutes later, in the same conversation, it argued for the opposite. Fluently. With the calm of something that had never held the other view. It could still recite the fact we’d established. What it had lost was the road we walked to get there.

I’ve hit that moment enough times to stop calling it a glitch. It keeps pulling me toward a smaller, stranger question than the ones people usually ask about AI. Not whether it’s smart enough, not whether it’ll take my job. Just this: what am I actually doing with it, that makes losing the road register as a loss at all?

Most people never feel that loss, because most people aren’t doing the thing that produces it. They ask, they get an answer, they leave: fetch and go. In that mode the model is a very good vending machine, and a vending machine that forgets your last purchase costs you nothing. Nothing was being carried forward.

Some of the time, though, you’re not fetching. You’re thinking. The reasoning accumulates across the hours, one decision resting on the last, a trajectory taking shape that exists only because you and the model built it turn by turn. It’s the same interface, but a completely different activity. And in that activity the answer is almost beside the point. What you’re working across is the gap: a partner who has read more than any person alive and remembers none of the road the two of you just walked.

I got interested enough in that gap to build for it. A memory layer whose only job is to keep the reasoning from evaporating between sessions, so the why survives and not just the what. I can see the shape of it. I can’t see all the edges yet, and plenty of it is still unproven. But building it showed me the thing I’d been too annoyed to notice: the friction was the point. When the model is confidently wrong about something we’d settled, that’s what makes me check, hold ground, catch it. I used to read that checking as the tax on thinking with AI. It turned out to be the thinking. The answers it handed me were mostly disposable. What lasted was the trajectory, the one I had to keep correcting because it kept slipping loose.

It forgets with total confidence, which is somehow worse than forgetting apologetically. But the confident version is the useful one. A tool that quietly agreed with me would let my own settled errors ride. The one that reasons straight past the settled point makes me re-earn it, or notice it was never earned.

That cuts both ways, though, and the second way is on me. It’s easy to let the machine do the remembering. Hand it the context, trust the summary, stop holding the thread myself. Do that and the trajectory still drops, only now I’m the one who dropped it. The tool’s amnesia I can catch. My own, the kind I invite by leaning on the tool, I usually can’t. So it only stays thinking if I keep carrying my half of it, if I still bother to remember what we settled and why. The moment I hand that over too, I’m back at the vending machine and can’t tell.

Underneath the prompt tips and the copilot demos, that’s the part nobody’s naming. Thinking with AI is a skill. A muscle. You build it or you don’t, and almost nobody is building it, because we’ve poured everything into making the model stronger and almost nothing into making the person across from it sharper. We train the tool to remember and reason and hold context. We don’t train the human to catch it when it’s confidently wrong, or to hold a settled point against a fluent case for the opposite, or to stay the one who remembers where the road went. That’s the muscle. It’s learnable, and it’s almost entirely untaught.

I can’t promise most people will build it. Plenty won’t; the tool is comfortable enough to lean on that the leaning becomes the habit. But the ones who do get something no upgrade to the tool can hand them. They stay the one who holds the why, working next to a machine that holds everything except that. The real work lives in that gap, and it answers to a trained human.

So the people who’ll get the most out of AI won’t be the ones with the best prompts. They’ll be the ones who trained themselves to think alongside it without quietly letting it think for them. That capacity was always there to build. Whether you build it is what decides how much of your own thinking you keep.