MemCoRe: Recovering Evidence from Progressively Compressed Factual Knowledge for Agent Memory

MemCoRe proposes a compressed hierarchical memory system aimed at addressing the issues of redundancy and retrieval efficiency encountered by large language model agents when handling interaction history. This system compresses detailed interaction records into keywords and theme groups, retaining the information structure required for retrieval at each level through a hierarchical structure, enabling cross-level retrieval of target evidence. Experimental results show that MemCoRe outperforms existing state-of-the-art baseline methods in comprehensive tests.