Hierarchical Graph Memory Boosts LLM Agent Efficiency and Reasoning

Large Language Model (LLM) agents often struggle with long-term reasoning due to inefficient memory management, where accumulating historical data can lead to irrelevant context and high retrieval costs. Existing graph memory systems store information in a flat structure, requiring repeated, independent updates for related changes. To address these challenges, a new framework called HiGram has been developed. HiGram employs an evolving hierarchical graph memory that organizes information into a coarse-to-fine architecture, significantly reducing the amount of irrelevant data during retrieval. Furthermore, HiGram introduces MicroGraph-based path-level localization to identify relevant subgraphs and evidence paths before updates, and a coordinated rewriting method that jointly revises both intra-unit memory and inter-unit dependencies. This approach ensures valid dependency structures are maintained within localized evidence paths, leading to substantial improvements in answer quality and token efficiency for long-term conversational Q&A and conflict-aware memory tasks.