EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective

Researchers introduced EvoMemBench, a unified benchmark for evaluating memory in LLM agents.

It measures memory along two axes: scope and content, covering in-episode and cross-episode memory needs.

Across 15 memory methods, results show that gains are task-dependent and no single approach consistently beats strong long-context baselines.

The benchmark helps separate memory evaluation from reasoning and planning, clarifying which memory designs fit different tasks.