Knowledge tracing is an important research area in personalized education, aiming to predict students’ future performance based on their historical interaction data. Existing knowledge tracing models face limitations in spatio-temporal information modeling. In the spatial dimension, the dynamic interactions between students and questions, along with peer effects among students, remain underexplored; In the temporal dimension, many existing works often ignore the modeling of student’s learning rhythm and the spatial structure hidden in temporal information. To address the above issues, this work focuses on the realistic scenario of streaming data, provides an incremental Spatio-Temporal Fusion Streaming Knowledge Tracing model, referred to as STSKT. Specifically, we first incrementally construct a dynamic heterogeneous graph to capture complex spatial relationships, including dynamical student-question interaction and peer effects. We then enhance temporal modeling from two aspects: proposing a rhythm-aware representation for learning rhythms, and capturing the hidden spatial structure by transforming recent interaction sequences into structured graphs. Finally, a three-channel fusion strategy is designed to capture the local relevance, global disparity, and holistic information for prediction. Experiments on four real-world datasets show that STSKT outperforms ten baselines, achieving improvements of 8.52% in AUC and 12.80% in ACC. Further analysis demonstrates that our key designs in graph construction, temporal modeling, and fusion mechanism effectively enhance prediction performance.