研究人员提出了Zero-Mem系统，旨在为大语言模型Agent提供无需消耗Token的内存操作机制[1]。该系统通过实体-上下文图和时间层级结构两种互补方式组织交互记录，在不调用大语言模型的情况下完成内存的检索和处理工作，仅在最终问答阶段才调用大语言模型进行推理[1]。
与现有基准方法相比，Zero-Mem将内存操作的时间成本降低了57.6%，同时维持了具有竞争力的性能表现[1]。该论文已于2026年7月31日提交[1]。
Researchers have introduced Zero-Mem, a system designed to enable memory operations for large language model agents without consuming any tokens from the LLM itself. [1] The system organizes interaction records through two complementary approaches: entity-context graphs and temporal hierarchies, allowing memory retrieval and processing to occur independently of LLM invocations. [1] By eliminating LLM calls during memory operations, Zero-Mem reduces the time cost of memory management by 57.6% compared to the fastest baseline method while maintaining competitive performance levels. [1]
The Zero-Mem framework defers LLM invocation until the final question-answering stage, streamlining the overall agent workflow. [1] This architectural innovation addresses a key inefficiency in current LLM agent systems, where memory operations traditionally require token consumption and repeated model calls. [1] The research was submitted on July 31, 2026. [1]