Last week a writing tool I built kept resurrecting a fact I had already killed.
The number was 231. Earlier in the conversation it had been 235, and the project had carried a working name I'd since replaced. I corrected both. The assistant agreed, apologized, fixed the draft. Three exchanges later it wrote 235 again, under the old name.
The easy read is that the machine forgot. That's not what happened, and the difference turns out to be the whole thing.
The correction was outvoted
When I went looking at the plumbing, the mechanism was embarrassingly plain.
The system was handing the model the last twenty messages of our conversation. Sitting in that window were a dozen of its own earlier drafts, each one confidently containing 235 and the old name. Also in that window, exactly once, was me saying "it's 231."
Nothing in the setup said which of those was true. So the model did what it always does, which is weigh what's in front of it. My correction was one voice against a dozen copies of a mistake, most of them written in its own hand. It didn't ignore me. It averaged me.
That reframing changed what I built. A model that forgets needs a bigger memory. A model that averages needs a hierarchy, and hierarchy is a thing you have to declare.
Where judgment gets stored
The fix wasn't a cleverer instruction. It was two structural moves, both boring, both about location.
First: the current file is the only truth about the article. Earlier drafts in the conversation are obsolete the moment the file changes. That rule is stated flatly, every turn, and the file is inlined every turn so there's no scenario where the assistant is reasoning from a stale copy because the fresh one wasn't handy.
Second, and this is the one I'd underrated: corrections get written down somewhere permanent. Every article now carries a brief, and the brief has a running section called Facts and corrections. When I fix something, the fix is recorded there in the same breath. That file goes into context on every single turn, forever, at full weight.
The correction stopped being a message and became a fact about the project.
Messages decay. That's not a defect, it's what a conversation is. Anything you say in a chat window has a half-life measured in turns, after which it's outside the window and functionally never happened. Anything written into a canonical document has no half-life at all.
I had been putting my most valuable input into the most perishable container available.
This is not a machine problem
Here's why I don't think this is a story about AI tooling, and why I've been chewing on it all week.
A correction is the smallest unit of judgment there is. It's you looking at something plausible and saying no, not that, this. It's a whole act of taste compressed into one line. It's also, almost always, the part of the work that only you can do, because the thing being corrected was competent enough to survive everyone else's attention.
And most organizations store those in conversation.
You've been in the meeting where a decision gets made properly, with good reasoning, and six weeks later the same decision is being made again from scratch because it lived in someone's memory of a Tuesday. You've watched a note in a document review get applied to that paragraph and to nothing else, so the same problem walks back in on the next draft. Nobody forgot, exactly. The correction just had nowhere durable to live, so it decayed on schedule.
The pattern is identical: judgment produced expensively, stored casually, re-derived at full cost every time. If you make the same correction three times, you didn't have a memory problem. You had a filing problem, and you paid for it in the only currency that matters, which is the attention you could have spent noticing something new.
Accumulation is not context
One more thing fell out of this, and it surprised me.
The article writer that has been quietly working for months, publishing to my sites every week, has none of this machinery. No conversation history at all. It reads a voice config and a brief, then drafts the entire piece in one shot.
It's better than the interactive version was. Not because it's smarter, but because it never accumulates. Nothing it produced earlier can compete with what I actually asked for, because nothing it produced earlier is in the room.
Long conversations feel like context and mostly aren't. They're sediment. Every turn adds material with equal standing to your intent, and after twenty of them your original instruction is one voice among many, gradually outvoted by the drift it was meant to prevent. Starting clean from a good brief beats twenty turns of correction, most of the time, by a wide margin.
The takeaway
Drafting is free now. What's scarce is the noticing, and a correction is noticing at its most concentrated.
So the question worth asking about any system you work inside, human or machine, is not whether it produces good output. It's whether your judgment survives being expressed. Does a thing you noticed once stay noticed? Or does it have to be renoticed on a schedule, at your expense, until you get tired and stop bothering?
Build the place where corrections are permanent before you build anything else. It's not a memory feature. It's the only thing that turns your attention into something that compounds.