Researchers introduce HORMA, a Hierarchical Organize-and-Retrieve Memory Agent that organizes working memory into a file-system-like hierarchy to reduce token usage and latency in long-horizon tasks.

HORMA (Hierarchical Organize-and-Retrieve Memory Agent) addresses the challenges LLM agents face in long-horizon tasks, such as context overload and loss of temporal structure. It structures the agent's working memory into a file-system-like workspace where raw interaction trajectories are organized into semantically structured, linked notes using file-system operations. A lightweight retrieval policy trained via reinforcement learning then navigates this hierarchy to extract minimal sufficient context for the current task. Across benchmarks like ALFWorld, LoCoMo, and LongMemEval, HORMA demonstrates superior efficiency-performance trade-offs, reducing token consumption in long conversations to as low as 22% of baseline usage.