ReTree: Self-Correcting Tree-Structured Memory for LLM Search Agents

A recent preprint on arXiv (2608.10676) presents ReTree, a tree-structured memory system designed for self-correction in search agents utilizing large language models (LLMs). This research tackles the challenges of limitless context expansion and noise when supplying complete execution paths to the LLM. Current compression techniques often sacrifice crucial details and fail to rectify incorrect information, hindering downstream reasoning. ReTree creates a limited reasoning context for each step while maintaining evidence linked to sources. It conceptualizes search as an evidence tree, with nodes that contain concise summaries, evidence, and revision histories. When new evidence contradicts prior claims, ReTree backtracks to the relevant node, updates evidence, regenerates summaries, prunes branches, and continues the search, enhancing the accuracy and efficiency of multi-step question answering in external contexts.