You finish a long session with an assistant understanding the problem better than you did going in. Then somebody else needs that understanding, you reach for the thing you would send them, and there isn't one. There is a transcript, and the transcript is not it.
That gap is not a missing feature. Context, in the sense that matters here, is not an object sitting somewhere without an export button. It is accumulated state, lossy by the time you go looking, compiled against things that live in you.
Context is state, not a document
The word invites you to picture a file. What accumulated over an hour of exchange is three kinds of thing, only one of them quotable.
Corrections. You said "no, the constraint is the review deadline, not the budget," and everything after was shaped by it. In the record that correction is one line in the middle; its effect is spread across everything downstream. A reader sees the line, not the reshaping.
Abandonments. You followed a direction for ten exchanges and dropped it. It is still load-bearing — it is why that option is off the table. But nothing announces "closed, and here is why"; the thread just stops. Silence reads as omission rather than decision, so a reader either re-raises it or assumes you never considered it.
Inferences nobody typed. From how you framed the problem, the model concluded you cared more about reversing the decision later than about moving quickly, and shaped its suggestions around that. You never said it. It may be right. You may not have noticed it happened.
The third is the sharpest. The load-bearing parts of a session are often the parts nobody wrote down — so they are the parts you cannot check, correct, or pass on without first noticing they exist.
The record is lossy too
Set the unwritten parts aside, and the copyable half is still not complete. Long sessions get compacted to fit a working window; early exchanges survive as a compressed summary or not at all, and outputs get truncated. What remains is a lossy compression.
So "just send the transcript" fails on its own terms before anyone reads a word. The practical consequences are a separate read. The point here is narrower: a record of an exchange and the state that exchange produced are different objects, and only one was ever copyable.
Context is indexed to you
Suppose the record were perfect — nothing compacted, every inference written down. It still would not transfer: it was assembled against knowledge not in it.
- Your priorities. The tie-break between two acceptable answers — never stated, because you never had to. The model absorbed it from what you pushed back on.
- Your risk tolerance. The same option reads as "fine" or "unacceptable" depending on what you can afford to be wrong about.
- Constraints that go without saying. The person who cannot be asked. The deadline not in writing. The approach that was tried before and went badly.
- Standing knowledge. You read "the migration" and know which one. Every piece of shorthand resolves against facts you carry with you.
A session is a conversation with someone who already holds all of that. Hand it to a person holding a different set and it does not decode — not because it is confusing, but because it was addressed to somebody else. The same happens to your own notes after enough years: the words survive, the index has drifted.
The consequence: there is nothing to export
Handing someone your context is not a transfer problem waiting on better tooling. There is no object to transfer.
What you can do is compose something for them, from scratch, that reconstructs enough state for them to answer. That is a different act from exporting, and it is work — which is why it usually does not happen. Naming it correctly changes what you reach for: you stop hunting for a share button and start writing a context brief.
What composing for someone requires
- State the inferences as inferences. If the reasoning assumed reversibility mattered more than speed, write it down and mark it a guess. Labelled, the reader can disagree — the most valuable thing they can hand back. Unlabelled, it passes as fact.
- Say what is closed, and why. One line per abandoned direction — the cheapest item here and the most reliably skipped.
- Re-index the shorthand. Every "the migration," "the client," "the thing we talked about" gets expanded or cut. If you cannot expand it, the reader could not have resolved it.
- Say what you want back. Context serves an ask. Without one, the reader cannot tell how much to care about, and defaults to none.
That produces a document that did not exist before — the tell that this was composition, not extraction.
An assistant can do it — check its first step hardest
An AI assistant is a reasonable author here: it holds the session state, it can render that state for a reader who has none of it, and it does not get bored on step three. That is what a briefing is — composed for a recipient, not excerpted from a session.
The place to look is step one. The inferences it made are, by definition, the parts never stated, so they are the likeliest to be rendered confidently and wrongly. Everything else is checkable against the record; that part is only checkable against you.
None of this is unique to AI. Transferring understanding always meant composing for the person receiving it. A long session just makes the failure easy to miss: it leaves behind a copyable artifact that looks like a complete record in a way memory does not.
If a session is open with a question that belongs to somebody else, the next step is composition, not export. Connect an assistant and have it draft the first version for you to correct.