ShardMemo

Official code for our EMNLP 2026 Findings paper. ShardMemo is a three-tier memory service for LLM agents in which metadata predicates first decide which shards a request may access, and a learned masked router then probes only a few of them (LoCoMo, HotpotQA, ToolBench).EMNLP 2026 Findings 论文的官方代码。ShardMemo 是面向大模型智能体的三层记忆服务，先由元数据谓词确定请求可以访问哪些分片，再由学习到的掩码路由器从中只探测少数几个（LoCoMo、HotpotQA、ToolBench）。