There is a sentence that has quietly become the standard promise of assistant software: it remembers you. Usually what this means, when you look at the implementation, is that it remembers what you typed into it. Those are not the same claim, and the distance between them is most of what this essay is about.
A chat history is a record of an interface. It knows what you said to it, in the order you said it, in the sessions you happened to open. It knows nothing about the eleven days you did not open the app, which may have been the eleven days that mattered. It knows the questions you asked, which are a strange and unrepresentative slice of a person — closer to a search history than to a life.
A life, if you had to say what it is made of, is people and feelings arranged in time. Not topics. Not sessions. Not queries. The people recur, the feelings move, and the arrangement in time is what turns the two into a story rather than a pile.
What a transcript is bad at
Consider a concrete case. Over four months you mention your sister eleven times. Twice in passing, warmly. Once at length, after an argument. Three times in a stretch of a fortnight where you keep circling something without naming it. Then five weeks of nothing, then once more, flat and brief.
The transcript holds all eleven. It holds them scattered across nine sessions, interleaved with everything else you happened to be writing about, in an order determined by when you had a free evening. Nothing in the transcript's structure says these eleven belong together. To see the shape you would have to read the whole thing and hold it in your head, which is precisely the thing you cannot do — it is why you wrote it down.
The thing that makes those eleven mentions legible is not more storage. It is a different organising principle: filed under your sister, in order, with the gap visible as a gap.
This is not an exotic idea. It is how human social memory already works. Daniel Wegner's account of transactive memory describes how people in close relationships distribute what they know across each other, and retrieve it by knowing who holds what rather than by searching their own store.1 We index each other by person. It is our native filing system, and no assistant that files by conversation is using it.
Two clocks, and why both matter
There is a second thing a transcript flattens, and it is subtler.
When you write something down, there are two moments involved: when the thing happened, and when you wrote about it. A chat log only records the second. It stamps everything with the moment you typed it, which is the moment least likely to be the interesting one.
You write about the argument four days later, once you have calmed down enough. You write about your grandmother the week after the funeral, not during it. You write about something from your twenties on an ordinary Wednesday in your thirties, because a song reminded you. In every case the timestamp on the message is a fact about your evening, not about the event.
Keeping both clocks separate is not a nicety. Endel Tulving's distinction between episodic and semantic memory turns on exactly this: episodic memory is memory located in time, tagged with a when, and that tagging is what lets you re-experience rather than merely know.2 Strip the when and an episode degrades into a fact. My grandmother died is a fact. My grandmother died in the October when everything else was also going wrong is a memory, and it means something the fact does not.
The asymmetry of what gets said
There is a third problem with treating chat history as life history, and it is the one that should make anyone cautious about the promise.
What you say to an assistant is shaped by what you think it is for. People ask assistants for things. They do not, mostly, tell them what happened. So the record skews hard toward the transactional and away from the ordinary, and the ordinary is where most of a life is.
Anything built on that record inherits the skew, and inherits it invisibly. It will be confident about the parts of you that were convenient to type into a text box and silent about everything else — and it has no way to know which is which. Ulric Neisser's study of John Dean's Watergate testimony is the canonical demonstration of how this goes wrong: Dean's recall was vivid, detailed, sincere, and systematically reorganised around his own role, in ways he could not have detected from inside.3 A record is not neutral just because it is accurate about what was said.
The honest position is that any written record is partial, and the useful thing is to know how it is partial. A record organised by person makes its own gaps visible — you can see that someone has not appeared in eight weeks. A record organised by session hides its gaps inside the scroll.
What we mean by the difference
So: most assistants remember your chats. We are trying to build something that remembers your life. The distinction is not rhetorical, and it comes down to four unglamorous properties.
It files by person, not by session. The eleven mentions of your sister live together, in order, whether they arrived in one evening or across four months.
It keeps two clocks. When it happened and when you wrote it are stored separately, because they are different facts and the gap between them is often the point.
Facts can stop being true. Something you said in March can be superseded by something you said in July without either being deleted. What changed, and roughly when, survives.
Absence is data. A person going quiet in the record is a fact about the record, and it is visible rather than buried.
None of that requires the system to interpret you, and it does not. It arranges. Once a week it shows you the arrangement. What the arrangement means is not its business — that part is yours, and a record that tried to do it for you would be worth less, not more.
If any of this resonated, you can read more about why people are the unit of a life or why journaling about people matters more than journaling about events.
- Wegner, D. M. (1987). Transactive memory: A contemporary analysis of the group mind. In B. Mullen & G. R. Goethals (Eds.), Theories of Group Behavior. New York: Springer. See also Wegner, D. M. (1995). A computer network model of human transactive memory. Social Cognition, 13(3), 319–339.
- Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of Memory. New York: Academic Press. See also Tulving, E. (1985). Memory and consciousness. Canadian Psychology, 26(1), 1–11, on autonoetic consciousness.
- Neisser, U. (1981). John Dean's memory: A case study. Cognition, 9(1), 1–22.
- Conway, M. A., & Pleydell-Pearce, C. W. (2000). The construction of autobiographical memories in the self-memory system. Psychological Review, 107(2), 261–288.