When your AI chat session hits its limit and starts a new one, what should it remember? A new arXiv paper tries to answer that with math instead of guesswork.
The paper treats "handover" - passing task context from one AI session to the next - as a formal transfer problem. Sessions end for a few reasons: the conversation exceeds the model's context window, the app restarts, or a different AI agent picks up the task. The authors propose a three-part record for what to carry forward: exact decisions and constraints, summary statistics justified by the task, and raw observations that those statistics would otherwise lose. They also work out, mathematically, how much memory a handover needs. For simple linear regression the answer is exact and finite. For more complex nonparametric cases, they give upper and lower bounds tying memory size to prediction error.
This matters because every AI product with a context limit - which is all of them - currently solves this problem with ad hoc summarization prompts, engineers guessing at what to keep. That approach has no way to know if it's losing information the next session actually needed. A formal minimum-bits answer means builders of long-running agents, coding assistants, and customer-support bots could eventually test their memory systems against a real benchmark instead of vibes.
It is worth noting this is theory, not a shipped tool - the paper offers bounds and a record format, not benchmarks against today's summarization techniques or a drop-in library. Whether real-world memory systems get anywhere close to these theoretical limits is the next question, and probably the more interesting one.