The session onboarding
Opening a new thread with just enough of yesterday.
The moment a user opens a new thread, two bad things can happen. The product can start with no memory, forcing the user to brief the model from scratch. Or the product can dump everything it knows into the thread, which feels invasive and rarely helpful. The middle path is a three-line recap, pinned to the top, dismissible in one tap.
The recap is the cliff-edge note from last session, shown as the opening line of this one. It carries enough to continue, not so much to confuse. The user reads, decides, and moves on in three seconds.
"The first message in a new thread shouldn't be blank. It also shouldn't be yesterday's noise."
Three lines. Carry, ignore, edit.
Three lines, not a paragraph. Two action items, one open question, or one status plus one blocker. The header says where it came from: the title of the previous thread, when it closed. Three buttons: carry (load this into the new thread's context), ignore (discard), edit (refine before carrying). Whatever the user does, the product learns the shape of their context switch.
- 01You were reviewing the Q3 plan draft; ACV targets were unresolved.
- 02Action item: confirm platform hiring backfill with Priya.
- 03Open question: do I keep the 500-seat floor or soften it?
Three lines. Dismissible. Editable. The first message in a new thread isn't blank, and isn't invasive.
Context is a gradient, not a switch.
Users switch sessions for two reasons: a new topic, or a clean slate on an old topic. The first wants no carry-over. The second wants some. A blunt toggle misses the middle case, which is most cases. A three-line recap with an edit option covers all three.
Recaps that help, not haunt.
- Three lines, hard cap. Longer recaps become the thing the user has to skim past. Three lines reads as reference, not noise.
- Name the source. The recap header says which thread it came from. Users can jump back if they need more.
- Edit → re-learn. When the user edits the recap, save the edits back. The next session's recap is shaped by the last session's corrections.
Recaps that carry too much.
The failure mode is a recap that includes everything: every decision, every open thread, every stale fact. The user drowns in history. Worse, the model quietly inherits stale state and answers current questions with last week's assumptions.
What this pattern gets wrong when it gets wrong.
- Stale memory
- A persistent fact about the user that's out of date and silently poisoning answers.
- Leaky context
- Content from another source, session, or user surfacing in a place it shouldn't.
- Empty disclaimer
- A legal-feeling warning that carries no specific information about this particular response.
Three shipping variants worth copying.
- A session-start card with three-line recap of last thread
- A toggle: carry, ignore, edit
- Edits are saved back to the carry-over for future sessions