一句话总结：这篇文章揭示了当前 LLM 智能体在处理长期、复杂交互任务时记忆能力的不足，并提供了一个更具挑战性的评估平台来推动该领域的进步 已有benchmark的特点 Large language model (LLM) agents have two complementary core capabilities: the ability to memorize task-relevant

kaixin li github repo: NVlabs/RocketKV: RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression To learn LLM KV Cache Compression October2001/Awesome-KV-Cache-