This study focuses on knowledge tracing, which involves monitoring the progression of learners’ competency levels and forecasting their future question-answering performance through examination of historical response patterns. To enhance the performance of existing deep learning models, this paper proposes a multi-feature fusion knowledge tracing model (BFKT) based on BERT and a forgetting mechanism. BFKT combines the basic cognitive level of students and the difficulty of the concept as new learning features, which are then encoded through the BERT layer. The model also introduces a custom-designed forgetting mechanism to simulate the forgetting process during the learning process of students. Experiments with the proposed BFKT model were conducted on three widely used datasets: ASSIST2009, ASSIST2012, and ASSIST2017, employing AUC (Area Under the ROC Curve) and F1 score as evaluation metrics. The results show that BFKT outperformed all RNN-based models on both metrics, demonstrating its ability to better capture students’ learning patterns and knowledge evolution. This highlights the effectiveness of its multi-feature fusion approach and the integration of the forgetting mechanism.