Knowledge tracing is an important research area in educational data mining. It focuses on using students’ historical interactions to dynamically model students’ knowledge state and predict their performance in future exercises. Traditional methods usually only analyze at the knowledge concept level without considering the difficulty differences between questions, which leads to prediction errors. Although some methods introduce question IDs to distinguish different questions, they still cannot fully reflect the relationship between question difficulty and students’ true level. To address this challenge, we propose a new difficulty-aware knowledge tracing model based on KAN. On the one hand, KAN’s unique edge-level learnable activation function is used to effectively capture the correlation characteristics and special error patterns between different questions in the interaction sequence; on the other hand, the correct rate of questions is converted into difficulty values based on the IRT framework, and the difficulty features are used to independently consider students’ memory strength and application ability to update the knowledge state. Extensive experiments on three public datasets show that our proposed KAN-DAKT model outperforms existing baseline models, especially in distinguishing question difficulty and accurately evaluating students’ true knowledge level.