Current research in knowledge tracing primarily focuses on predicting students’ knowledge states, but existing methods still have limitations in feature utilization and temporal dependency processing. Addressing the limitations of the DGEKT model in feature representation and temporal modeling, this paper proposes a Deep Graph Knowledge Tracing model incorporating Self-Attention and Forgetting mechanisms (Self-Attention Forgetting Graph Knowledge Tracing, SAFKT). The model achieves accurate knowledge state tracking through the collaborative work of three core modules: First, the knowledge state encoder enhances feature representation capability through a multi-layer perceptron (MLP), optimizing feature transformation with layer normalization and ReLU activation functions; Second, the self-attention mechanism captures complex relationships between knowledge points through a multi-head attention mechanism, calculating the correlation between features using scaled dot-product attention; Finally, the forgetting mechanism simulates the forgetting characteristics in human learning processes through the collaborative action of time decay gates and knowledge state gates, controlling forgetting intensity using sigmoid functions and exponential decay functions. These three modules are organically integrated through a graph neural network structure, where the output of the knowledge state encoder serves as input to the self-attention mechanism, and the features processed by the self-attention mechanism are further adjusted by the forgetting mechanism, ultimately forming a complete knowledge state representation.