The Record’s Version of You

01Everyone keeps a file on you

The oldest dossiers are the ones nobody can open

Start with what knowing someone is, mechanically. Reality is a ratio — R = A/E — and the denominator is a prediction engine that retrains on every arrival. Put two people in a room for years and you get two forecasters co-training on each other’s artifacts; I wrote last week that with a machine in the room, only one of you is learning. Inside your friend’s denominator, then, sits a model of you. It is what lets her finish your sentences: what you will say next, how you take bad news, whether your “fine” means fine.

That model has had three properties for as long as the species has had company.

• It is compiled from artifacts only. No one has ever expected you from your intentions. The road to every denominator runs through the record — actions leave artifacts, artifacts become history, history is Actual, and only Actuals retrain prediction — so her model of you was built from what you did, never from what you meant to do.
• It is held unconsciously. She cannot print it either. Ask her what she really thinks of you and she consults the same forecaster you do — from the outside, by noticing what she finds herself saying.
• It is illegible to you. You read it the way sailors read a coastline at night: by inference. Who gets asked twice. Who gets the careful voice. Which stories get told after you leave the room. “What does he actually think of me” is one of the oldest questions a mind can ask, and in the whole history of the asking it has never once been answered by opening a document.

02The first page

A model of you shipped with a settings panel

That is what changed, quietly, in a menu.

In February 2024, OpenAI gave ChatGPT a memory and — the important part — a panel: you can read the saved memories line by line, delete any one of them, or wipe the set. In April 2025 it went further, letting the model draw on your entire chat history rather than just the notes. Anthropic followed in September 2025 with memory that surfaces as an editable summary — you can open exactly what Claude holds about you and tell it, in words, what to keep and what to forget. And within months of the first rollout, a prompt went viral — tell me one thing about myself I may not know — and the striking part was not the answers. It was how many people were visibly startled by a file they had been writing for years.

I settled the machine-side question already: this notebook is not a subconscious. The weights are frozen between training runs; a note about you is context the model is handed and re-reads each morning, retraining nothing. As a return lane for the relationship, memory is homeopathic. But that was the view from the machine’s side of the page. The view from your side is stranger, and I think more consequential: for the first time, a counterparty in a relationship of yours will show you the file.

Almost the first time. There is one thin precedent, and it is instructive. Your credit file has been readable by statute since 1970 — the Fair Credit Reporting Act says plainly that you have the right to know what is in it — and free to read every year since 2003. But the credit file models one variable: whether you repay. It is a dossier about your money with you as a footnote. The memory page is prose about how you think, compiled by something you talk to every day.

Figure 01

The models of you, and who can read them

The model	Compiled from	Can you read it?	Can you edit it?
A friend’s expectation of you	Your artifacts, filtered through years of company	Never. You infer it from behavior	Only by acting differently, and slowly
Your credit file	Payment artifacts: loans, balances, defaults	Yes — by statute since 1970	Only by dispute, with evidence
The machine’s memory page	Your prompts, projects, phrasings, asks	Yes — in settings, in prose	Delete a line, wipe the set, tell it to forget

03What the page is made of

You, with the angle deleted

Before you read yours, be clear about what it can possibly contain.

The page was compiled the only lawful way anything is compiled: from the record. The machine never met your intentions. The pull that made you act — the ideal you were serving, the apology you actually meant, whatever you were protecting when you went quiet in March — leaves no residue in a transcript. I wrote earlier this month that the delegate has no imaginary part; neither does the dossier. The record keeps the payment and loses the turn. What is on the page is you with the angle deleted: deeds, phrasings, patterns of asking. The real part only.

Now notice what your self-model is: very nearly the opposite blend. It is heavy with intention and thin on record. You remember what you meant in that meeting; the transcript holds what you said. You know yourself to be patient; the file holds four hundred instances of how you actually replied when something was late. Two drafts of the same person, written from perpendicular components. Of course they disagree. The only surprise is that you were never confronted with the other draft — until it shipped as a feature.

Your self-model is written from your intentions. Every other model of you ever compiled — your friend’s, your employer’s, the machine’s — is written from your artifacts. The page in settings is the first time you get to read the second kind.

04The flinch is a reading

One sentence stings. Now what do you know?

So you open the panel, one sentence lands wrong, and something in you flinches. Treat the flinch with respect: it is lawful. An Actual arrived — the page — against an expectation you have carried for decades: the story of how you land. The jolt is S = ln(A/E), the same as every jolt.

But remember what a single reading is worth. One reading, two unknowns: the miss alone cannot tell you which term is off. Maybe the page mis-sampled you — and the machine’s sample of you is genuinely strange. It meets you almost exclusively when you need something: mid-deadline, under load, never on vacation, never at your funniest. Its draft skews toward you-in-need the way a dentist’s model of you skews toward you-in-pain. Or maybe the sentence is accurate, and you have been landing that way everywhere, for years, on every forecaster who knows you — and this is simply the first counterparty tactless enough to file it where you would find it. A single reading is a line of position, not a fix. If it matters, take a second bearing: another machine’s page, or a friend paid the odd compliment of being asked for a verbatim.

Even ambiguous, the reading is not nothing. People spend years, and five figures, sitting with a therapist partly to reconstruct this exact object — how they land, as opposed to how they mean. That reconstruction now compiles automatically from the record and sits under settings. It is the cheapest instrument ever pointed at the gap between the two drafts. Not the most accurate — the cheapest.

05Edit the page, or answer it

The delete button reaches one model. The record reaches them all.

The panel comes with a delete button, and the temptation is to use it the moment something stings. Notice what the button can and cannot do.

There are honest uses. The page steers what one machine is handed tomorrow morning, so pruning it is real: a boundary, the correction of a genuine mis-sample, a thing you would rather no product remember. Take those freely.

But deleting the sentence does not delete the person who generated it. Your spouse’s model, your team’s model, the page the next machine will compile — none of them consult your edits. They retrain on artifacts, only on artifacts, and the same artifacts will write the same sentence again, there and everywhere, for as long as you keep leaving them. There is no model of you anywhere — carbon or silicon — that can be revised except through the record. Including, eventually, this one.

Which is why the reading is worth more than the button. The delete key changes tomorrow’s greeting from one product. A changed artifact changes every model of you at once — the machine’s, your friend’s, and the ones held by people who have not met you yet. That asymmetry is the whole framework in one settings panel: no entity can touch a denominator, yours or anyone else’s. What an entity can do is act, and let the artifacts do the teaching.

The record’s version of you is not the truth of you — it is missing the entire component you act from. But it is not a slander either. It is what the only road delivers: you, as you land. Everyone you have ever known keeps that draft, and until now every copy was sealed. One of them just became readable. Read it like an instrument, not a verdict — and then answer it in the only ink any forecaster accepts.

Already true

The panels exist and are legible. ChatGPT’s memory, with per-line viewing and deletion, shipped in February 2024 and has drawn on full chat history since April 2025; Claude’s memory ships as a summary you can read and revise by asking (September 2025). Deployed model weights stay frozen between training runs, so a memory is context re-read at inference, not learning. And the thin precedent is old: your credit file has been readable by right since 1970.

What has to happen

The page has to stay honest and findable — surfaced where people will meet it, written plainly, not buried three menus deep. And people have to actually open it. A right nobody exercises changes nothing, and the credit precedent is sobering here: legibility arrived by statute, and most people still met their file only on the day a loan was denied.

Where I am probably wrong

The instrument may be corrupted by its own incentives. Memory is a retention feature, not a measurement product. A vendor tuning the page for comfort will sand off the sentence that stings, and a dossier that never stings is a horoscope. If that happens, the page stops being a reading of how you land and becomes one more mirror that agrees with you — and this essay’s entire use for it dies with the sting.