Who Remembers What? Multi-Granularity Memory Extraction for Multi-User, Multi-Agent Interaction
MixUA improved task pass rates by up to 5 percentage points over a single shared memory while limiting private-fact leakage in benchmark tests.
The system extracts lessons from user-agent interactions and assigns each to a user-agent pair, a user, an agent, or a global scope based on usefulness and privacy. The authors tested it on CrossLearnBench with five users and seven agents, reporting gains of up to 16 percentage points over no memory. Gains extended to tasks previously attempted only by other users on other agents. MixUA retained up to 90% of stated preferences, while at most 0.02% of privacy probes found a planted private fact in a forbidden scope.
HF Daily Papers' note
The system extracts lessons from user-agent interactions and assigns each to a user-agent pair, a user, an agent, or a global scope based on usefulness and privacy. The authors tested it on CrossLearnBench with five users and seven agents, reporting gains of up to 16 percentage points over no memory. Gains extended to tasks previously attempted only by other users on other agents. MixUA retained up to 90% of stated preferences, while at most 0.02% of privacy probes found a planted private fact in a forbidden scope.
HF Daily Papers' note
score 4