ReTree Improves LLM Search Agents with Self-Correcting Tree Memory

Large language model (LLM)-based search agents often struggle with long-horizon tasks due to unbounded context growth and the propagation of errors when providing full execution trajectories. Existing context compression methods frequently sacrifice important details or replace erroneous facts without correcting subsequent reasoning. To overcome these limitations, researchers have introduced ReTree, a novel self-correcting tree-structured memory mechanism. ReTree manages search as an evidence tree, where each node stores bounded summaries, evidence, and revision histories. This design allows it to maintain a manageable per-step reasoning context while preserving the provenance of evidence. A key feature is its ability to self-correct: if newly retrieved evidence contradicts a previous claim, ReTree traces back to the origin of the claim, updates the outdated evidence, regenerates summaries, prunes affected branches, and then resumes the search process. This source-grounded evidence provenance ensures reliable conflict localization and keeps final claims traceable to their original passages. Experiments on four public question-answering and search benchmarks demonstrate that ReTree consistently outperforms traditional full-trajectory ReAct methods, improving answer accuracy by up to 25.6 percentage points while significantly reducing the average maximum per-step reasoning context.