Graph Augmented Tree Search

Graph-Augmented Tree Search (GATS) is a planning framework for long-horizon AI agents that combines tree search (systematic look-ahead over possible actions) with graph-structured memory (structured state and relationship tracking). The core idea is to use a layered world-model approach: a high-level planner proposes candidate actions and decompositions, while lower-level checks refine and validate choices against constraints and observed outcomes. This OSF project hosts the preprint, supporting materials, and links to the implementation. Code: https://github.com/MMWilliams/gats Canonical preprint/DOI: https://doi.org/10.5281/zenodo.18264145 https://doi.org/10.17605/OSF.IO/U7EJK Keywords: agent planning, tree search, world models, MCTS/UCB, graph memory, long-horizon agents, tool use