Recommender systems have become essential infrastructure for mitigating information overload in online platforms, where modeling fine-grained user interaction sequences is critical for personalization. With the increasing scale of user data, the need for efficient algorithms to process and model large interaction sequences has grown. While sliding window mechanisms are commonly used to capture evolving preferences, their simplistic design often overlooks critical behavioral anchors and temporal-behavioral semantics, thereby yielding passive and noisy representations that undermine robust sequential modeling. To address this, an efficient sliding window-based model for temporal action modeling (SWinAct) is proposed. Specifically, this model applies behavior-time aware augmentation jointly constrained by temporal intervals and behavioral semantics to generate semantically consistent augmented samples. In addition, a multi-anchor guided sliding window strategy is designed to dynamically center on key behaviors with predefined sliding ratios, thereby preserving decision coherence. A boundary handling strategy and a short sequence protection strategy are also developed to prevent excessive truncation or indiscriminate processing. Finally, extensive experiments on three real-world datasets validate the superiority of SwinAct over existing state-of-the-art methods.