EdgeMem Builds LLM-Free Agent Memory from Evidence-Preserving Hypergraphs

Summary

EdgeMem is an agent-memory method designed to retain answer-bearing details from earlier interactions without repeatedly asking a generative LLM to summarize them. The authors argue that summaries and other generated representations can discard details before the system knows what a future query will require. EdgeMem preserves the original interaction turns and organizes them with complementary content, temporal, and episodic anchors in a hypergraph built through lightweight local processing. Its retrieval stage returns source evidence directly, leaving LLM use for final answer generation. On the LoCoMo and LongMemEval-S benchmarks, the method shows strong retrieval and memory-grounded question-answering performance. On LoCoMo, it records a strict-judge score of 61.01, the highest among seven reproduced systems evaluated with a shared prompt, compared with 58.70 for the next reported result. Both memory construction and retrieval require no generative-LLM calls. The results support preserving and structurally organizing original evidence as an efficient foundation for multi-session agent memory, while the abstract does not provide further implementation or error analysis.