As an important technology for intelligent education, the core task of knowledge tracing is to track changes in a student's knowledge level, which is manifested in predicting whether the student can answer the next exercise correctly. Memory augmentation networks excel at this task due to the idea of external memory. However, existing memory augmentation networks only utilize a small number of features of interaction records, such as an exercise’s tag and response. A Knowledge Tracing Model based on Multiple Behavior Features Embedded Memory Networks (MFKT) is proposed to fully utilize the learning and forgetting features in interaction records. The model considers both learning and forgetting behaviors in the learning process. Firstly, two major features, learning and forgetting, are extracted from the interaction records, and then the extracted learning features are embedded into the memory network by scalar crossover, while the forgetting features are embedded by vector combination, which is used to enhance its learning ability for the students' interaction sequences. In addition, this paper also considers the difference in knowledge level growth after different students' answers are completed and adds a knowledge growth layer to the original memory network for calculating the knowledge growth obtained from students' responses. Experiments on public datasets show that MFKT is more in line with the real learning patterns of students and can realize more accurate tracing of students' knowledge status.