Human behavior is deeply influenced by emotions, yet large-scale analysis of emotion-behavior dynamics remains challenging. We propose EMG-Mamba, an emotion-aware multimodal framework that integrates GraphSAGE for spatial-social relations with Mamba for temporal dynamics. Our mood embedding module infers users’ emotional states from check-in data using spatial diversity, activity intensity, and sentiment-weighted POI categories. Experiments on real-world datasets show EMG-Mamba outperforms baselines with 10% improvement in Acc@5, while maintaining high computational efficiency. The model also uncovers psychologically-grounded emotion-behavior patterns, validating its interpretability. This work enables scalable emotion-aware trajectory analysis for behavior modeling and mental health applications.