Agent memory enables LLM agents to leverage earlier interactions when answering new queries. Existing approaches usually compress interaction histories into summaries or other LLM‑generated representations; repeated generation incurs cost and may discard answer‑bearing details before the system knows what a future query will need. EdgeMem follows a simple principle: keep the original conversation turns and organize them with complementary content, temporal and episodic cues. In practice EdgeMem builds a multi‑anchor hypergraph by lightweight local processing, where each anchor captures a content, time or episode feature and hyperedges link related turns. Retrieval returns the source evidence directly, reserving LLM use for the final answer generation, thus providing structured access to multi‑session histories while faithfully preserving the original dialogue. Experiments on LoCoMo and LongMemEval‑S show strong retrieval and memory‑grounded QA performance. On LoCoMo, under a shared prompt EdgeMem achieves the highest strict‑judge score of 61.01, surpassing the best of seven reproduced systems at 58.70, and both construction and retrieval require no generative‑LLM calls. Overall the results demonstrate that preserving and organizing source evidence offers an effective and efficient foundation for agent memory without generative memory management.
Review