Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite
HiGram localizes the evidence path before rewriting agent memory.
The paper proposes a hierarchical graph memory that separates coarse upper-level nodes from finer MemoryUnits to reduce irrelevant retrieval context. Its MicroGraph step identifies a support subgraph and evidence path conditioned on the query and update. The rewrite then updates both memory contents and dependencies along that localized path. The authors report better answer quality, token efficiency, and conflict-aware evidence selection than baselines. ArXiv · AI/CL/LG's note
The paper proposes a hierarchical graph memory that separates coarse upper-level nodes from finer MemoryUnits to reduce irrelevant retrieval context. Its MicroGraph step identifies a support subgraph and evidence path conditioned on the query and update. The rewrite then updates both memory contents and dependencies along that localized path. The authors report better answer quality, token efficiency, and conflict-aware evidence selection than baselines. ArXiv · AI/CL/LG's note
score 5