A three-part record for what to write down when an agent session hits the context wall

arXiv 2608.14528·medium signalXBlueskyLinkedInCopy link

This paper models compaction and session restart as transferring an in-context learning state, and separates exact material recovery from preserving the target distribution — two goals most summarizers conflate. It proposes a three-part handover record: decisions and constraints stored verbatim, task-justified statistics for repeated evidence, and the raw original observations whose effect the statistics do not capture. Theory is developed through Gaussian linear regression (an exact finite-dimensional handover exists) and nonparametric regression (upper and lower bounds tying the memory budget to squared prediction error), and it isolates the cost of writing the record before you know the downstream query.