The CAS team proposed the hypergraph memory architecture HyperMem, achieving a breakthrough in long-conversation AI memory capabilities

ME News Report, April 11 (UTC+8), a research team from the Institute of Information Engineering, Chinese Academy of Sciences, published a paper at ACL 2026 main conference, proposing a hypergraph memory architecture called HyperMem for long-term dialogue. It achieved an evaluation accuracy of 92.73% on the LoCoMo long dialogue benchmark, setting a new state-of-the-art level. Existing RAG and graph-structured memory schemes often rely on pairwise relationships, making it difficult to capture high-order associations among multiple elements, resulting in fragmented retrieval content. HyperMem introduces a hypergraph structure that divides dialogue memory into three levels: topics, segments, and facts, and uses hyperedges to unify related segments and facts. Coupled with hybrid lexical-semantic indexing and a coarse-to-fine retrieval strategy, it enables precise and efficient retrieval of high-order associations. Experimental results show that on the LoCoMo long dialogue benchmark, HyperMem achieved an accuracy of 92.73% in the LLM-as-a-judge evaluation, reaching the current best performance (SOTA), validating its effectiveness in maintaining consistency and personalization in long-term dialogues. This research offers new ideas for dialogue agents to maintain contextual coherence, track ongoing tasks, and provide personalized services during long-term interactions. The paper has been published on arXiv (ID 2604.08256). (Source: BlockBeats)