Researchers have developed a new hierarchical memory system called Weighted Memory Tree (WMT) to help large language model agents remember relevant information over long periods. Unlike existing approaches that focus on storing more data, WMT assigns dynamic retention scores to each memory, allowing it to selectively retain useful information and suppress irrelevant content. The authors evaluate WMT on several benchmarks, showing a 9.97% average accuracy improvement and 32.8% reduction in prompt-token usage compared to linear memory. They also demonstrate that WMT limits the persistence of unreliable information. --- Why it matters: This matters because long-horizon language model agents often struggle with managing their execution histories, which can lead to decreased reasoning quality due to outdated or misleading information. WMT's ability to selectively retain useful information and suppress irrelevant content is crucial for improving the accuracy and efficiency of these models. Source: https://arxiv.org/abs/2608.20631