Our direction, while Bubbling AI is in development. Examples below illustrate the experience we want to build, not a claim that every view is available today.
A conversation can start with a clear purpose and become harder to follow as it grows. Which decision still stands? Did the assistant carry forward the correction? Is a useful detail saved for next time, or only present in this conversation? What is still waiting on you?
We think those are product questions. The person should not have to reconstruct the whole discussion to work out where things stand.
Three things that should be easier to distinguish
Saved context is information intended to remain useful across work: notes, documents, preferences, and memory records. You need ways to inspect and correct it.
The current task needs a more focused working set: the goal, relevant material, constraints, decisions, and unresolved questions. A saved fact is not necessarily relevant to every task.
The current state of the work is what has happened, what is underway, what needs your input, and what comes next. A long chat transcript is not always the clearest way to communicate that state.
We are building around these distinctions. Selected references, saved memory, progress, and requests for your input should make the work easier to steer. The assistant still selects context; we are not claiming a complete display of everything a model considered.
A useful handover should be concrete
Imagine returning to a proposal after a few days. A useful working summary might say:
We are preparing a proposal for a small team. We agreed to keep the first version simple. The audience is confirmed; pricing is still open. The next step is an outline using the brief you selected.
That is a small example, but it gives you something to inspect. You can correct the audience, revisit the earlier decision, or answer the open question. The discussion can continue from an understandable state.
A summary can also be wrong or incomplete. It needs a way back to the relevant sources and a way for you to correct it. Presenting a confident paragraph without those connections would not solve the underlying problem.
Focus and efficiency should help each other
Model processing has a real cost. Sending everything repeatedly can waste time and attention; stripping too much away can remove the detail that makes an answer useful.
Our aim is relevant context that supports the task without unnecessary repetition. We want context management to improve both the quality of the work and the resources needed to do it. That is a design goal to test, not a claim that the shortest prompt always wins or that every task costs less.
We care about whether the assistant respects the constraints, produces a useful result, and needs fewer corrections. Token counts alone cannot tell us that.
Clear context does not replace enforced boundaries
A reminder in a prompt is useful, but important permissions cannot depend on a model remembering every instruction.
The surrounding system needs to check access, applicable approvals, and other required conditions at the point where an action occurs. A model should not be able to skip those checks by reporting that it completed them. These controls do not guarantee perfect judgment or obedience on every ordinary task; they establish specific boundaries outside the model's discretion.
The result we want is a discussion partner that can stay focused while you retain a clear sense of the work: what matters now, what has been decided, and where your judgment is needed next.
Read how we design around changing models and why working knowledge deserves privacy.