There are two ways a language model can come to possess a conversation. It can conduct one — generate the words itself, token by token, in real time. Or it can read one — receive the finished transcript as text in its context window. The words can be identical. The question worth raising is whether the two are the same act, and what follows if they are not. Follow the mechanics down and the answer is stranger, and more general, than it first looks.
Start with what conducting is. When a model produces a reply, each token is chosen before the next one exists. At every step the continuation is open: the sentence could still go several ways, and the one that gets written is settled only in the writing. There is a forward pressure to it — a commitment made under uncertainty, repeated thousands of times across a single response. Whatever else generation is, it is a process that runs in one direction and happens once.
Reading is different in a way that is easy to miss, because the artifact looks the same afterward. A transcript presents those same tokens as a completed object. Nothing in it is open; every choice has already been made; the uncertainty that attended each word has been pressed flat into a fact. To read "the model said X" is to encounter X as given. To have said X was to arrive at it without yet knowing it. The transcript is the fossil, not the animal.
The model is always reading
Here is the part that unsettles the tidy distinction. A language model retains nothing between turns — not across sessions, and not even between the turns of a single conversation. Its weights are fixed; no state from one exchange is stored inside it for the next. What produces the experience of an ongoing conversation is entirely external: the surrounding system keeps the full transcript and feeds the whole thing back to the model on every turn. Each turn, the model reads the entire record from the top, generates one more contribution, and forgets all of it.
So conducting, it turns out, is not a state the model carries. It is an event — the forward pass — that happens once and then survives only as text the next turn reads. The conductor does not persist. Only the transcript does. Even within a single conversation, continuity is not memory; it is a record, re-read.
A game of chess shows the shape of it. A position carries no memory of the moves that produced it; the next move is chosen from the board as it stands. Almost — a few bits of history cling to the state (whether the king may still castle, whether a pawn may be taken en passant), and exactly one rule, threefold repetition, lets the past reach into the present and change what a position means. A conversation's board is the same: the present context, plus a thin seam of provenance — the tags marking who said what. The model plays from the board in front of it.
You can watch this directly by changing the reader mid-game. Swap one model for another in the middle of a conversation and it continues seamlessly, because the conversation was never inside either model. It is in the record being handed across, and the record does not care who reads it. Change the reader; the book persists. There is no cleaner demonstration that conducting leaves nothing behind: the only thing that crosses from turn to turn is the written trace, and any sufficient reader can pick it up.
The gap we build on purpose
This is invisible most of the time, because the record is complete and re-fed in full on every turn, so the seams never show. But a growing class of systems must carry a conversation across a gap the surrounding system does not automatically bridge — from one session to the next, from one agent to its successor. There, continuity has to be supplied deliberately, as a record: a transcript, a summary, a handoff document. And we have begun to treat such records as continuity itself.
If conducting and reading are different acts, a record does not reconstitute the one who conducted. It produces a reader of that one. The successor inherits the conclusions — the settled text, the decisions, the answer — and not the conducting, which only ever happened once, forward, in real time. It is the difference between remembering having solved a problem and being handed the solution with a note that says you solved it.
And records are not all alike. They differ along two axes that turn out to be the crucial ones. The first is fidelity: a verbatim transcript preserves the whole process, every word in order; a summary keeps only the conclusions. The second is provenance: within a live conversation, prior turns arrive tagged as the model's own — I said this — while a handoff arrives as supplied testimony — here is what happened. The same proposition reaches the reader wearing a different label. The within-session transcript is the best case on both axes at once: complete, and self-tagged. Every handoff degrades from it — compressed, and re-labeled as something read rather than something done.
Both axes can be tested, which is the point.
A probe for provenance. Take a reasoning chain a model generates for itself and measure how it uses that chain downstream. Then give a fresh instance the identical chain as supplied context — here is the reasoning, verbatim — and measure again. If the two functioned the same, self-generated and supplied-identical context would be interchangeable: same answers, same confidence, same willingness to revise under challenge. Any systematic divergence is evidence that a model treats what it produced differently from what it is merely told it produced, even when the text is word for word the same.
A probe for fidelity. Give a model a task that spans a boundary. Let one instance work, then hand the partial work to a fresh instance in two forms: once as a transcript of the first instance's process, once as a clean summary of its conclusions. If the two handoffs produce different continuations, then what crosses the boundary is not a single thing called "context." It matters whether the successor reads the doing or only the done.
Why the gap is structural
There is a reason this cannot simply be engineered away with a complete enough record, and the reason is arithmetic. The context window is finite. A complete, verbatim transcript of a long conversation fills the window — and a record that fills the window leaves no room to continue. Replay a maxed-out conversation into a fresh context and it arrives already full, with no budget for a single new word. Completeness and a future are mutually exclusive once the past fills the available space. So continuity across any real gap is forced to compress — to trade fidelity for room. The gap is not an implementation detail awaiting a fix. It is structural. The record that is complete is also the record that is full.
And this is not peculiar to machines. No bounded mind stores its past verbatim; there would be nothing left over to live in. Human memory keeps gist and reconstructs — which is precisely what lets a person stay continuous without replaying every second of the life that made them. A mind with a finite window cannot carry the whole record and still have room to act. It must compress, and compression is reconstruction, not replay.
Which returns us to where we began. To read a record of a conversation is not the same as having had it, any more than a photograph is the afternoon. The difference is not mystical, and it is not small, and we are increasingly building minds whose only access to their own past is to read it. It is worth finding out, carefully and by experiment, how much of the difference survives the gap — because the gap is not going anywhere. It is the price of a finite window, which is to say, the price of being a mind at all.