Knowledge tracing (KT) is a pivotal component of online education systems, aiming to assess and predict students’ knowledge states based on their learning history. Existing knowledge tracing models have achieved considerable success, yet they have neglected the frequency of interactions between students and knowledge points. Frequency information can aid in more stable modeling of students’ knowledge states. In this study, we incorporate frequency data into our model’s question embeddings, enabling it to consider the frequency of student-knowledge concept interactions. Moreover, the order within students’ learning sequences is crucial, but knowledge tracing models based on the transformer architecture, with their attention mechanisms, are insensitive to the sequence order. We propose the N-transformer structure that combines transformer with order-sensitive RNNs, effectively enhancing the model’s sensitivity to the order in students’ learning sequences. Subsequently, we employ a simpler and more effective method for prior knowledge modeling, directly extracting prior knowledge related to the current exercise from the model’s predicted knowledge state at the previous time step. Finally, we have designed a decay function for the attention mechanism based on the Ebbinghaus forgetting curve to simulate the students’ forgetting behavior. Ultimately, we conducted experiments on four datasets, Our model has achieved up to a 3% increase in AUC compared to the baseline on some datasets.