Weighted Memory Tree Enhances LLM Agent Performance

The Weighted Memory Tree (WMT) is a novel hierarchical memory system designed to improve the performance of large language model (LLM) agents, particularly in complex, multi-step tasks. As LLM agents execute long-horizon tasks, their growing histories can become costly and introduce irrelevant or outdated information, degrading reasoning quality. WMT addresses this by organizing execution into tasks, subtasks, and actions, assigning a dynamic retention score to each memory. WMT continuously updates these scores based on events and applies selection-based decay, allowing it to intelligently preserve critical information, fold completed trajectories, and suppress low-utility content. This ensures that the agent's active memory remains focused and relevant. Evaluations on the GAIA-Text benchmark, using models like Qwen3-8B and Llama-3.1-8B, demonstrated significant improvements. WMT increased accuracy by nearly 10 percentage points on average and reduced prompt-token usage by over 30% compared to linear memory. Furthermore, memory-poisoning experiments showed WMT's effectiveness in limiting the propagation of unreliable information, highlighting its robustness.