Large language model (LLM) agents can solve multi‑step tasks that require planning, tool use, and external information access. As execution histories grow, inference cost rises and the reasoning process becomes exposed to outdated, irrelevant, or misleading information, which can degrade quality. Existing memory approaches merely organize or compress histories and lack a mechanism to decide which memories stay active. We introduce the Weighted Memory Tree (WMT), a hierarchical memory system that structures execution into tasks, subtasks, and actions, assigning each memory a dynamic retention score. Event‑driven updates and selection‑based decay adjust these scores, allowing WMT to preserve useful information, fold completed trajectories, suppress low‑utility content, and still access folded context. We evaluate WMT on the GAIA‑Text benchmark using Qwen3‑8B, Gemma 4E4B, and Llama‑3.1‑8B, conducting ablations and memory‑poisoning experiments. Compared with linear memory, WMT improves accuracy by about 9.97 percentage points while reducing prompt‑token usage by 32.8%. Poisoning experiments show that WMT limits the persistence and propagation of unreliable information. The results suggest that effective long‑horizon agent memory relies more on selectively activating information than on storing more data.
Blogger's Review: The Weighted Memory Tree’s dynamic retention strategy keeps essential context while cutting computational overhead, offering a compelling direction for making long‑horizon LLM agents more practical.