Wang et al. introduce MixUA, a memory system for multi-user, multi-agent environments that distills independent user-agent interaction trajectories into reusable lessons and assigns each lesson a scope—user-agent pair, user, agent, or global—based on utility and privacy constraints. To evaluate this approach, the authors construct CrossLearnBench, a benchmark combining established agent tasks with simulated user preferences and privacy probes designed to detect leakage of planted private facts into unauthorized memory scopes. Experiments involving five distinct users and seven heterogeneous agents demonstrate that MixUA improves task pass rates by up to 16 percentage points compared to a no-memory baseline and by 5 points over a single global memory pool, including on tasks previously attempted only by different users on different agents. The system retains up to 90% of previously stated user preferences across sessions. Crucially, privacy leakage remains minimal: at most 0.02% of privacy probes recover a planted private fact in a forbidden scope. These results indicate that granular, scope-aware memory extraction enables effective cross-agent and cross-user knowledge transfer while maintaining strict data isolation, offering a scalable architecture for co-evolving user-agent ecosystems.

Read original