Temporal knowledge graph reasoning aims to infer implicit knowledge from historical snapshots. Effective reasoning generally requires both local and global information. Traditional methods mainly focus on local modeling of several recent snapshots to capture evolutionary patterns, such as structural dependencies and temporal evolution, but they usually fail to exploit global semantics and achieve adaptive integration. Existing approaches attempt to incorporate both local and global contexts, yet still face two challenges: limited ability to capture latent associations across time and insufficient fusion of multi-level representations. To overcome these limitations, we propose EPGLA, a novel reasoning framework that integrates an Evolutionary Graph Unit, an Association Graph Unit, and a Gating Unit. EPGLA explicitly builds association graphs across different timestamps to uncover hidden relations and global semantics, while a gating mechanism adaptively combines local and global representations. Extensive experiments on five benchmark datasets demonstrate that EPGLA consistently outperforms state-of-the-art extrapolation models, confirming its effectiveness in temporal knowledge graph reasoning tasks.