Substantially written by GPT-5.6 Sol and edited by Allan.
Somewhere in the last year, a lot of us quietly changed how we think.
Not what we think about — how. The AI became a place to work things out. You explain a half-formed idea to it and hear it come back sharpened. You argue with it and find out which of your reasons were real. You use it as a mirror, or a foil, or a patient colleague who will read the same messy draft eleven times without sighing. Whatever you call this, it isn't looking things up. It's thinking, conducted out loud, with a partner.
And if you've done it seriously for more than a few months, you've had the sinking feeling too.
You're working on something — a business plan, a research question, a book, a decision you can't stop turning over — and you realize you have been here before. You worked this exact problem through. You reached a conclusion. You remember the feeling of having settled it. You cannot remember what you settled, and you cannot find the conversation where you settled it.
So you do it again. A little worse than last time, because last time you had reasons and now you only have a conclusion.
The problem isn't memory. It's that nothing accumulates.#
The obvious diagnosis is that the AI forgot. That's not it, or not the important part. The problem is that you have no place to stand.
Think about what you actually own after six months of this. You own a pile of conversations. Hundreds of them, maybe thousands, in one app or several, with titles the machine invented, on whatever device you happened to be holding. Each one was useful when it happened. Together they are a heap.
And if you think this doesn't apply to you because you use one assistant and stick to it — check. You've been asking Google questions all year and Gemini has been answering them. Your phone has an assistant. Your work laptop has a different one. Meanwhile the people getting serious about this are deliberately spreading out, because the models have genuinely different temperaments and are good at different things, and because handing one model's output to another is one of the more productive things you can do with them. That's worth trying. It also doubles the problem, which is my point: the more useful this gets, the more places your thinking ends up, and none of them are yours.
A heap is not a body of work. Nothing in it points at anything else. Nothing says this superseded that. Nothing records that in March you thought the answer was X and by June you'd changed your mind, and why. The good sentence you wrote at 11pm in a conversation about something else is in there somewhere, and it will not be found, and effectively it does not exist.
I've started thinking of it as a documentary problem. You've shot hundreds of hours of footage. Every hour of it was worth shooting. But no one is cutting a film, and film is the only reason anyone shoots footage. What you have is an archive of raw material and no work.
Here's the part that makes this worse than it sounds. Depth is cumulative. Anything worth thinking hard about — a real design, a real argument, a real strategic question — takes more than one sitting. It takes many sittings that each start where the last one ended. If every sitting starts from roughly zero, you don't get depth no matter how many sittings you do. You get breadth, forever. Which is exactly what most people report: I use it constantly, it's amazing, and I'm not sure I'm getting anywhere.
Programmers solved this, and nobody noticed it was a general problem#
There is one place where people routinely do go deep with AI, and it's worth asking why.
Programmers work with these tools all day, on projects that run for months, with results that visibly compound. It's tempting to credit the models — they're unusually good at code. I think the real reason is duller and more important.
Programmers have somewhere to put the answer.
When a programmer finishes a work session, the outcome doesn't live in the conversation. It lands in the project itself, which is a single durable thing that holds the current state of the work and its entire history: what changed, when, why, what was tried and abandoned. Tomorrow's session opens that, not yesterday's chat. Nobody scrolls back through last month's discussion to recover a decision, because decisions were never stored in discussion. They were deposited into the work.
That's the whole trick. It isn't clever tooling and it has nothing to do with programming. It's just that the conversation is treated as how the work gets done, and something else entirely is treated as the work.
For everything else we think about — the strategy, the argument, the book, the decision — we have no equivalent. The conversation is the only artifact. So the conversation gets asked to be both the workshop and the warehouse, and it's a terrible warehouse.
The missing thing#
What's missing is small and boring to describe, which may be why nobody's built it.
For each thing you're actually developing, there should be one living document that is the state of your thinking. Not a transcript. Not notes you'd have to reread to reconstruct. A page or two: what you now believe and why, what you've decided and what convinced you, what's still open, what you tried and rejected — so you never re-litigate it by accident. Every claim on it linked back to the conversation where it came from, so you can always go see the footage behind the cut.
Then the loop changes shape. A session opens by loading that document, so you begin where you left off instead of near the beginning. And a session ends by updating it. That's the missing gesture. Right now conversations just stop. They should close the way a good meeting closes: here's what changed, here's what we decided, here's what's still open.
Back to the documentary for a moment, because the division of labor there is instructive. A producer knows what film they want to make and never watches the raw footage. A director owns the story. An editor goes through everything and proposes the cut. You want to be the producer — you say what the work is for, and you judge results. The machine can do the other two jobs: one part of it holding the through-line, another part grinding through the material and coming back with "here's what I'd keep, here's the order, here's why." You look at the proposed cut and say yes, or no, not that. What you should never have to do is watch the footage again. That's the job you keep accidentally giving yourself now.
And once your thinking sits in one place, something becomes possible that today can't happen at all: the machine can notice when you contradict yourself. In March you concluded one thing; in June, working from a different angle in a different app, you concluded the opposite, and both conclusions are sitting there with the reasoning behind each. That's not an error to be silently smoothed over — that's the single most interesting event in a long piece of thinking. It should be raised, with both sides shown, and you should be the one to decide which of you was right. Nothing in the current setup can even see it. The March conversation and the June conversation have never met.
You don't get to keep the old ground. That's the whole complaint. Your thinking should be somewhere that isn't scattered across six apps and four devices and a thousand conversations you'll never open again.
Why now, and why not them#
Fair objection: people have been pushing personal knowledge systems for decades — notebooks, wikis, card files, second brains. They mostly fail, and they fail for one reason. Somebody has to do the filing. After a hard hour of thinking, nobody wants a second unpaid job tidying up what they just did. So the system decays, and the decay is the norm, not the exception.
What's genuinely new is that the filing can now be done for you. Summarizing a session, extracting what was decided, noticing that today's conclusion contradicts March's — this is precisely the kind of work these models are good at. Your job shrinks to reading a short proposed update and saying yes, or no, that's not what I meant. That's a few seconds. That's sustainable. The chore that killed every previous attempt is the one thing the new tools handle easily.
So why hasn't it been built?
Because the labs have decided chat isn't the frontier anymore. The frontier is work: agents that run for hours, do the research, write the code, produce the deliverable. Enterprise budgets, white-collar tasks, measurable output. Watch where the new surfaces are going and it's unmistakable — the conversation is being demoted to a control panel for something else.
Here's the thing: they're making the same observation I am. They've noticed that the conversation isn't the point either, that the valuable thing is what comes out the other end and persists. We simply disagree about what comes out. They think it's the spreadsheet, the report, the completed ticket — the artifacts a company pays for. I think that for anyone using these tools to work out what they actually believe, the thing that should persist is the thinking itself. They want to cash out in the enterprise. I want to cash out in my intellectual life. Nobody is building the second one, because it isn't where the money obviously is.
And there's a quieter reason worth naming without melodrama. A durable, portable record of your thinking is, structurally, the thing that would make it easy to leave. Every provider now has several apps that don't share history with each other, exports that arrive days late, and work that's stuck on the device where you started it. You needn't assume malice; the incentives are clear enough on their own. Nobody is racing to build you a warehouse they don't own.
Which means, for now, it's on us — and it starts with a change of habit more than a change of tools. Decide what you're actually working on. Keep one page for each. End sessions by updating it. Do it by hand if you have to; ask the machine to draft the update, which it will do well.
But the larger point is the one I'd like to leave you with. These tools are the most interesting thinking instrument to arrive in a very long time, and we are using them in the least cumulative way imaginable — brilliant sessions, evaporating on contact. The bottleneck was never how well the machine thinks. It's that nothing is being built out of all that thinking.
Shoot less footage. Cut a film.