Modeling dynamic graphs in continuous time is critical for applications such as user behavior prediction and recommendation systems. These models can effectively capture fine-grained and long-term temporal dependencies. However, existing approaches often suffer from high computational costs and optimization difficulties, especially when handling time-sorted neighborhood sequences over long horizons. In this work, we propose DyG-Hyena, a novel continuous-time dynamic graph learning framework that combines conditional variational autoencoder (CVAE)-assisted temporal modeling with efficient feature fusion. Our approach has two main innovations: (i) Efficient temporal fusion—we replace the Transformer with an improved, lightweight Hyena module to model and fuse time-sorted neighborhood feature sequences, reducing computation of this process while maintaining accuracy. A CVAE layer is added before Hyena to capture relative time constraints, enhancing generalization for link prediction. (ii) Task-specific multi-dimensional information enhancement—for link prediction, we incorporate cross-order neighborhood intersection encoding; for node classification, we introduce statistical encoding of node features. Extensive experiments on benchmark dynamic graph datasets demonstrate that DyG-Hyena achieves excellent performance while substantially reducing temporal modeling complexity. Our code is available at https://github.com/yangchang666/DyG-Hyena.