Next Point-of-Interest (POI) recommendation is a key service in location-based social networks (LBSNs) that leverages users’ trajectory data to provide personalized location recommendations. Its goal is to uncover users’ latent preferences and recommend appealing locations for their next visits. Although previous studies have effectively captured movement patterns and users’ latent intentions within trajectories, there still exhibit certain limitations in modeling global POI transition relationships and uncovering users’ temporal preferences across different time periods. We introduce a model for next POI recommendation that is based on Adaptive Graph and Time Tree (AGT2) in this study, which dynamically captures the POI graph topology through the adaptive graph, providing more expressive POI embedding. And then we introduce a novel data structure, the time tree, to organize users’ check-in records across multiple time granularity levels. This structure effectively captures user preferences over different time scales. By propagating and aggregating information beginning at the lowest level and advancing to the top level, the time tree enables comprehensive, multi-level analyses of user behavior, providing a nuanced understanding of temporal patterns. In conclusion, in order to assess the effectiveness and robustness of AGT2, we perform comprehensive experiments on three datasets.