CAS Team Proposes Hypergraph Memory Architecture HyperMem, Long Dialogue AI Memory Capability Achieves Breakthrough

BlockBeats News, April 11th. A research team from the Institute of Information Engineering, Chinese Academy of Sciences, presented a paper at the main conference of ACL 2026, proposing the HyperGraph Memory Architecture for Long Conversation, HyperMem, achieving a 92.73% evaluation accuracy on the LoCoMo long conversation benchmark, refreshing the current state-of-the-art level.

Existing RAG and graph-based memory solutions often rely on pairwise relationships, making it difficult to capture high-order correlations among multiple elements, leading to fragmented content retrieval. HyperMem introduces a hypergraph structure, dividing dialogue memory into three levels: topic, snippet, and fact. It uses hyper-edges to uniformly group related snippets and facts. Combined with a hybrid lexical-semantic index and a coarse-to-fine retrieval strategy, it achieves precise and efficient retrieval of high-order correlations.

Experimental results show that on the LoCoMo long conversation benchmark, HyperMem reached a 92.73% accuracy in the LLM-as-a-judge evaluation, reaching the current state-of-the-art (SOTA) level, validating its effectiveness in maintaining consistency and personalization in long conversations. This research provides new insights for dialogue agents to maintain context coherence, track persistent tasks, and provide personalized services in long-term interactions. The paper has been published on arXiv (ID 2604.08256).