Long-running agents depend on memory of their past interactions to keep track of task state—for example, what has already been done and which conditions remain true. The authors of a new arXiv paper argue that this state is usually derived: no record states it directly, so the agent must compute it from the data it has seen.
To study this problem, the paper introduces MemDerive, a diagnostic benchmark for agent memory over evolving data. The benchmark is designed to test how well agents can maintain and update derived state as new information arrives, rather than simply retrieving stored facts.
The work highlights a gap in current evaluation: many benchmarks focus on factual recall, but real-world agent tasks often require inference over a changing history. MemDerive aims to provide a targeted way to measure this capability and diagnose where agents fail.