EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective

In three linesEvoMemBench is a unified benchmark evaluating LLM agent memory across two axes: scope (in-episode vs. cross-episode) and content (knowledge-oriented vs. execution-oriented). Comparison of 15 memory methods: long-context baselines remain highly competitive, retrieval-based methods dominate knowledge-intensive tasks, procedural methods excel for execution-oriented tasks.

Summary generated by Claude — human-verified