paper web signal

MixUA scopes agent memory into four tiers, limits leaks to 0.02%

TL;DR

  • MixUA assigns each lesson a self-evolving agent learns to one of four scopes: a user-agent pair, a single user, a single agent, or global.
  • In tests with five users and seven agents, MixUA raised task pass rates by up to 16 points over no memory and 5 over a single global memory.
  • The system retained up to 90% of previously stated user preferences while at most 0.02% of privacy probes found a planted fact in a forbidden scope.

In a new arxiv preprint, researchers propose MixUA, a memory system that sorts every lesson a self-evolving AI agent learns from experience into one of four scopes (a specific user-agent pair, a single user, a single agent, or global) and report that the design raises task pass rates by up to 16 points over systems with no memory while limiting private-fact leakage to at most 0.02% of adversarial probes.

The paper, co-authored by Yunzhe Wang, Mesut Erhan Unal, Sheng Zhang, Shabnam Ghadar, and Peng Tang and dated November 7, frames shared agent memory as a policy question rather than a storage one. "Sharing across users and heterogeneous agents requires deciding who can benefit from each lesson and what information must remain private," the authors write, arguing that most group-evolving agent systems skip this step entirely.

Their evaluation uses a new benchmark, CrossLearnBench, which "pairs established agent tasks with user preferences and privacy probes that test whether private facts leak into memory other users can read." With five users and seven agents, MixUA also gained 5 points over a single global shared memory and retained up to 90% of stated user preferences, including on tasks that had previously only been attempted by other users on other agents.

The abstract names no baseline scoped-memory system to compare against, and the privacy probes were planted by the authors themselves rather than by an independent red team. "We hope MixUA offers a scalable path toward a global system where users and agents co-evolve," the paper concludes.