Multivariate time series forecasting (MTSF) is a fundamental task for modeling complex dynamic systems, with broad applications in areas such as weather prediction, energy management, finance, and healthcare. However, real-world multivariate time series often exhibit temporal distribution shifts and dynamically evolving inter-variable dependencies, which undermine the assumptions of static modeling frameworks. Moreover, most existing architectures treat temporal relations as symmetric, disregarding the causal flow and inherent directionality of sequential interactions, thereby limiting their ability to capture asymmetric spatio-temporal dependencies. To address these challenges, we propose DiSRet—a Directional Structure-aware Retentive Network that adaptively models evolving dependencies across time and variables. The proposed Directional Structural Preference Module (DSPM) learns directional sparse graphs by incorporating causal masking and learnable temporal biases, enabling the model to encode asymmetric and time-aware dependency structures. The Structure-Guided Retention Mechanism (SGRM) further integrates these structural preferences into the representation learning process, facilitating adaptive modulation of temporal, spatial, and spatio-temporal dependencies. Extensive experiments on seven real-world forecasting benchmarks demonstrate that DiSRet consistently outperforms state-of-the-art methods, achieving superior accuracy and strong adaptability to dynamic and asymmetric dependencies.