The Day My Own Brain Went Stale
I build a knowledge layer. My own knowledge went stale, and nothing caught it. This is what happened.
July 29. That is the last-edited stamp on the catalog page where the two hundred nuclei of this series live, and on that day the page was true. Engrym was two people. The catalog said so in several places, not as a headline but in passing, the way you mention the weather: a two-person company that runs like twenty, real numbers from a live two-person brain, two people and fourteen specialist agents. The white paper said it. The thesis pages said it. The vault that mirrors the thesis pages said it, because that is what a mirror does. Older drafts said it, because they were written from the same sources. None of those sentences was a decision. Nobody had ratified “Engrym has two founders.” It was simply the condition the company was in when the documents were written, and the documents absorbed it the way they absorb everything else.
Then the condition changed. I won’t narrate why, because it isn’t the point and it isn’t anyone’s business. By the end of the summer Engrym was one person, me, working with a swarm of agents, and I knew it the way you know your own address. What I did not do, on the day it became true, was open the catalog. Or the white paper. Or the vault. Nothing in the stack was touched. No page changed. Every mirror still said two.
The stack ran fine for weeks. This is the part worth sitting with. The agents kept drafting, the drafts kept arriving, the figures kept rendering, and nothing complained, because nothing had any way to complain. A page that has gone false looks exactly like a page that hasn’t.
A draft came back with a sentence in it. The sentence said “we,” and the “we” meant two humans, and a paragraph later it said “a two-person shop,” about me, in the present tense. The agent that wrote it had done nothing wrong. It had been told, in a standing instruction I wrote myself, to always use the thesis document as its source. It used the thesis document. The thesis document said two. The agent reported what the source said, with the fluency and confidence I pay for, and the sentence was finished, grammatical, and false.
I went looking for how far it had spread. The phrase was not in one place. It was in the catalog, three times at least. It was in the white paper’s framing sections. It was in the vault, in every file that mirrored a page that had it. It was in a bio. It was in product copy that an agent had generated from the bio. Each copy had been made in good faith by something, a person or an agent, that trusted the copy before it, and each trust was reasonable at the time. It was in the connective tissue of dozens of pages, never as a claim with an owner and a date, always as an aside inside a sentence about something else. That is the shape of most of what a company believes about itself: not filed, repeated.
The obvious thing to do was fix the documents, and I started. I stopped after the second page, for a reason I want to state carefully because it is the whole lesson. The phrase “two-person” also appears in the series as an example about someone else: a two-person shop rents the same intelligence as a two-thousand-person org. That sentence is not about me, was never about me, and is still true. A human reading the two usages tells them apart in half a second. A find-and-replace cannot, and neither can an agent working from the text alone, because the difference between a claim about Engrym and a figure of speech about a hypothetical shop is not in the words. It is in what the words refer to, and reference is not a property of the page.
So I did not fix the documents. I wrote a new one. It is short. It says: Engrym is built and run by one person, me, with a swarm of agents. It says that some sources, and it names them, contain a stale claim to the contrary, and that every such claim is superseded. It tells every agent that reads it to flag the stale claim once if it finds it in a source, then proceed with the corrected framing. It even covers the edge case, the rhetorical two-person shop, and says to leave that one alone. I ratified it by hand, which in my stack means I read it twice and put it where every session starts. It is, in every respect that matters, a supersession record. It just happens to be written in prose and filed as a memo, because at the moment I needed it, the stack had no slot for it.
I have reread that memo a few times since, and what strikes me is not that it exists. It is what it had to do. It had to tell the agents that the source they were told to trust was wrong about one thing, and to keep trusting it about everything else. It had to describe the scope of the wrongness in words, because there was no field to hold it. It had to anticipate the one case where the same words are correct. A status flag does all of that in a byte. I wrote a page.
The argument I would normally make here, I don’t need to make, because the memo made it. The claim “Engrym is two people” did not degrade. It did not drift. It went from true to false in a single day, on an event that happened outside every document, and no document knew. Nothing in a page knows it has gone false. The reconciliation I run against code would not have caught it, because the shape of the company is not in the repository. The retrieval layer would not have caught it, because the stale sentence was the most-repeated sentence on the subject and therefore, as the relevance article showed, the most retrievable one. The only trigger in the whole stack was a human reading a draft on a week when he had time to read drafts. If that week had been busier, the sentence would have shipped under my name.
That is why the product exists, and I want to be exact about the “why.” Not because staleness is a theoretical risk I identified from a distance. Because it happened to the person building the fix, in the knowledge base that describes the fix, while he was writing about the fix, and the only thing that stopped it was luck. If someone can show me a stack where that claim would have flipped itself, with no human reading a draft and no status held outside the text, I will retire the memo. Until then I am building the slot it should have had.
The catalog page still says July 29. The pages still say two. The banner is a memo.