EdgeMem addresses the limitations of existing agent memory techniques that rely on LLM summarization. The method focuses on maintaining the original interaction turns, organizing them through content, temporal, and episodic cues. This is achieved through the construction of a multi-anchor hypergraph using lightweight local processing. Retrieval directly returns source evidence, reserving LLM usage for final answer generation. This approach combines structured access to multi-session histories with faithful retention of the original conversation. Experiments on LoCoMo and LongMemEval-S show strong retrieval and memory-grounded question answering. Specifically, on LoCoMo, EdgeMem achieves the highest strict-judge score among seven reproduced systems (61.01) compared to 58.70. Construction and retrieval require no generative-LLM calls. This indicates a more efficient and cost-effective approach to agent memory.
Source: https://arxiv.org/abs/2609.05553