Temporal graphs play a crucial role in modeling evolving relationships across domains such as social networks, financial transactions, and healthcare systems. This research presents an innovative temporal graph learning framework that combines evolving node embeddings with a flexible attention mechanism to effectively model time-sensitive interactions. The proposed method demonstrates superior performance in terms of predictive accuracy and adaptability when compared to conventional techniques. Experimental results reveal a 15% increase in accuracy and a 30% reduction in computational overhead relative to baseline models such as recurrent neural networks and traditional graph neural networks. Moreover, the framework shows enhanced resilience to incomplete data, improving efficiency by 20% in sparse data scenarios. These advancements contribute to more reliable applications in event prediction, fraud detection, and anomaly recognition, ultimately supporting better decision-making across sectors like finance, healthcare, and cybersecurity.