Most large language model agents are tested in static, unchanging environments. This creates a dangerous illusion of competence. In the real world, conditions shift constantly, and current AI systems are not built to handle the chaos of evolving tasks.

Researchers have introduced EvoArena, a new benchmark suite designed to expose this fragility. Instead of static tests, EvoArena simulates dynamic environments where terminal commands, software states, and even social preferences change progressively over time. The goal is to see if agents can adapt their knowledge and behavior as the world around them updates.

The results are stark. Current agents struggle significantly, achieving an average accuracy of just 39.6% across these evolving domains. To fix this, the team developed EvoMem, a memory system that records updates as structured histories. By tracking how their memory changes, agents can reason about environmental evolution rather than just reacting to isolated prompts.

EvoMem consistently improves performance, boosting average scores on EvoArena by 1.5%. It also lifts performance on standard benchmarks like GAIA and LoCoMo by 6.1% and 4.8%, respectively. Crucially, it improves chain-level accuracy by 3.7%, helping agents complete sequences of related subtasks that require sustained adaptation. Mechanistic analysis reveals that EvoMem improves evidence capture in the memory, indicating better preservation of complete evolving environment states. This highlights a critical need for modeling evolution in both evaluation and memory architecture.