HORMA: Hierarchical Memory System Improves LLM Agent Performance on Long-Horizon Tasks
Researchers introduced HORMA, a hierarchical memory organization system designed to help large language model agents handle long-horizon tasks more efficiently by structuring experiences like a file system rather than relying on lossy compression or similarity-based retrieval. The system organizes memories into two stages—structured construction and navigation-based retrieval—using reinforcement learning to select minimal yet sufficient context. This approach matters because it addresses a fundamental limitation of LLM agents: their inability to maintain context over extended task sequences, reducing token usage by up to 78% while maintaining or improving task performance.
HORMA (Hierarchical Organize-and-Retrieve Memory Agent) tackles the problem of LLM agents struggling with long-horizon tasks due to their stateless nature, which forces all task-relevant information into growing input contexts that degrade reasoning quality and increase computational costs. The system organizes experience into a hierarchical, file-system-like structure where summarized entities link to corresponding raw trajectories, enabling efficient access without information loss. The approach uses two key mechanisms: a construction module that iteratively refines how experiences are structured by distinguishing between failures caused by missing information versus misleading or overloaded context, and a navigation module that uses a lightweight reinforcement-learning-trained agent to traverse the hierarchy and retrieve task-relevant context. Across three benchmark environments (ALFWorld, LoCoMo, and LongMemEval), HORMA achieved better efficiency-performance trade-offs than existing methods, using at most 22.17% of baseline token usage in long conversation tasks while improving task performance under constrained context budgets.
What's missing
The paper does not discuss potential limitations of the hierarchical organization approach, such as how the system performs when task structures are highly non-linear or when causal dependencies are ambiguous. Additionally, the generalization claims to 'unseen tasks' lack detail about the degree of task dissimilarity tested or how performance degrades as task distribution diverges from training data.
What different sources said
- arXiv cs.AICenter
Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.