EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments

EvoArena benchmarks LLM agents across terminal, software, and social domains where environments change progressively over time, exposing a significant performance gap versus static evaluations. The accompanying EvoMem memory system stores structured update histories as patches rather than snapshots, letting agents track how their knowledge has changed. Current agents struggle on EvoArena, suggesting static-environment benchmarks meaningfully overestimate real-world agent reliability.