Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents

The article introduces HORMA (Hierarchical Organize-and-Retrieve Memory Agent), which enhances long-horizon task performance in large language models by organizing experiences in a hierarchical structure akin to a file system. This approach allows for efficient navigation and retrieval of task-relevant information while maintaining detailed context, achieving up to 22.17% of baseline token usage in long conversations. HORMA's two-stage working memory system—structured memory construction and navigation-based retrieval—demonstrates improved efficiency and performance across benchmarks like ALFWorld, LoCoMo, and LongMemEval, making it a significant advancement for practitioners dealing with complex, multi-step tasks in AI development.