Abstract
TRACE
TRACE studies which memories a returning agent may act on after collaborators have changed the shared state. Its training-free admission layer compares the agent's departure checkpoint with updates made during its absence, resolves invalidated information, and assembles a bounded view that covers the returning role's remaining obligations. Evaluations examine whether agents can preserve useful memory while rejecting stale information in evolving multi-agent collaborations.
Temporal memory admission helps returning agents retain valid information and reject stale state as multi-agent collaborations evolve.