Recall or Recheck? Evaluating Memory-Augmented Agents in Evolving Mobile Environments

Long-term memory allows agents to retain past experience as their environments continue to change. Historical information may disappear from the current state, while recent updates may be absent from memory. We introduce $R^2$-MobileMemory, an executable benchmark that separates persistent memory from the query-time state of the same evolving mobile world to study when agents should recall, recheck, or combine both sources. Our evaluation shows that memory provides its largest gains when task-critical historical evidence is no longer available in the environment, whereas it adds little when current-state search is already sufficient. Extending memory with later observations does not consistently improve task success. The added information can overlap with evidence already available in current applications or introduce competing versions of relevant facts. Even after retrieving a complete annotated support set, agents can still fail to select the relevant state or realize the requested outcome. These results frame memory-augmented interaction as a problem of temporal evidence arbitration, where success depends not only on what memory preserves, but also on selecting and using the right evidence relative to an evolving environment.