HiGram: Hierarchical Graph Memory for LLM Agents
Researchers have introduced HiGram, a hierarchical graph memory framework designed to improve long-term reasoning in LLM agents. The system organizes memory into a coarse-to-fine architecture with upper-level nodes and MemoryUnits, reducing irrelevant information during retrieval. It also implements path-level localization and rewriting to efficiently update memory as new facts and feedback arrive. The approach addresses limitations of flat graph memory, which can accumulate historical data and increase retrieval costs. HiGram is detailed in a paper on arXiv (2608.05095), submitted in August 2026.
Key facts
- HiGram is a hierarchical graph memory framework for LLM agents.
- It uses a coarse-to-fine architecture with upper-level nodes and MemoryUnits.
- It reduces irrelevant information during retrieval.
- It supports path-level localization and rewriting.
- It addresses issues in flat graph memory.
- The paper is available on arXiv with ID 2608.05095.
- The paper was announced in August 2026.
- The framework is designed for long-term reasoning.
Entities
Institutions
- arXiv