Memory Flow-Controlled Knowledge Tracing with Three Stages
Academic Background
With the rapid development of artificial intelligence (AI) technology, intelligent tutoring systems (ITS) such as Khan Academy and Coursera have made significant progress in personalized learning. Knowledge Tracing (KT), as a key technology in ITS, aims to infer students’ knowledge mastery and predict their future learning performance by analyzing their learning data. Despite recent advancements in the field of KT, existing models have shortcomings in simulating memory structures, leading to inconsistencies between students’ explicit learning and implicit memory transformation. To address this issue, a research team from Central China Normal University and other institutions proposed a Memory Flow-controlled Knowledge Tracing model with Three Stages (MFcKT), which aims to improve the accuracy and interpretability of KT by simulating three stages of memory: sensory memory registration, short-term memory encoding, and long-term memory retrieval.
Source of the Paper
The research was conducted by Tao Huang, Junjie Hu, Huali Yang, and other scholars from the Faculty of Artificial Intelligence in Education at Central China Normal University. It was published in 2025 in the journal Neural Networks under the title Memory Flow-controlled Knowledge Tracing with Three Stages. The study received support from multiple institutions, including Central China Normal University, Ningxia Normal University, and Wuhan Textile University.
Research Process
1. Problem Definition and Theoretical Foundation
The study first defines the Knowledge Tracing (KT) task, which involves inferring students’ latent knowledge states and predicting their future performance based on sequences of learning data (including exercises, concepts, and responses). Based on Information Processing Theory, the research decomposes the learning process into three memory flow stages: sensory memory registration, short-term memory encoding, and long-term memory retrieval. These stages correspond to different types of memory (sensory, short-term, and long-term) and simulate the dynamic evolution of knowledge through memory transformation mechanisms.
2. Model Design
The core of the MFcKT model lies in simulating memory flow through three stages: - Sensory Memory Registration (SMR): Uses contrastive pre-training and self-attention mechanisms to extract sensory memory from learning sequences. - Short-term Memory Fusion (SMF): Fuses the relational and temporal features of sensory memory through a dual-channel structure (including attention mechanisms and recurrent neural networks) to generate short-term memory. - Long-term Memory Retrieval (LMR): Designs a monotonic gating mechanism to compute the weights of hidden memory states and perform read-write operations on the memory matrix, ultimately combining long-term and short-term memory vectors to retrieve latent knowledge states.
3. Experimental Design and Datasets
The study conducted extensive experiments on five public datasets (ASSISTments2009, ASSISTments2015, ASSISTments2017, Algebra2005, and NIPS34) to validate the effectiveness of MFcKT. The datasets cover student learning records from different intelligent tutoring platforms, including exercises, concepts, and response data.
4. Experimental Results
The results show that MFcKT significantly outperforms existing knowledge tracing models on all datasets. For example, on the ASSISTments2009 dataset, MFcKT achieved an AUC (Area Under Curve) of 0.8232, a 1.18% improvement over the state-of-the-art model. Additionally, MFcKT performed well on the ASSISTments2015 dataset, which lacks exercise information, with a 3.48% improvement in AUC, demonstrating the advantage of its dual-channel structure in capturing relational and temporal features.
5. Ablation Experiments
To verify the contributions of each module, the study conducted ablation experiments. The results show that removing any module (SMR, SMF, or LMR) leads to a decline in model performance, further proving the importance of the three-stage memory flow design.
Main Results and Conclusions
1. Sensory Memory Registration Module
Through contrastive pre-training and self-attention mechanisms, MFcKT effectively extracts sensory memory and captures individual differences among students. Experiments show that this module plays a key role in improving the model’s prediction accuracy.
2. Short-term Memory Fusion Module
The dual-channel structure (attention mechanisms and recurrent neural networks) successfully fuses the relational and temporal features of sensory memory, generating more representative short-term memory. This design significantly enhances the model’s ability to model the learning process.
3. Long-term Memory Retrieval Module
The monotonic gating mechanism computes the weights of hidden memory states, enabling effective management and retrieval of long-term memory. Experiments show that this module plays an important role in improving the model’s interpretability and prediction performance.
4. Comprehensive Performance
MFcKT outperforms all baseline models on the five datasets, demonstrating its superiority in the knowledge tracing task. Additionally, the model’s efficiency and robustness provide strong support for its application in real-world educational systems.
Research Highlights
- Innovative Memory Flow Design: MFcKT is the first to introduce the three-stage memory flow theory into the field of knowledge tracing, effectively addressing the inconsistency problem in memory modeling of existing models.
- Dual-channel Structure: By combining attention mechanisms and recurrent neural networks, MFcKT captures both relational and temporal features of memory, enhancing the model’s overall performance.
- Contrastive Pre-training Technology: Using contrastive pre-training, MFcKT better simulates individual differences among students, enhancing the model’s personalized learning capabilities.
- Extensive Experimental Validation: Experiments on multiple public datasets demonstrate the superiority and robustness of MFcKT, laying the foundation for its practical application.
Research Significance and Value
The research on MFcKT not only provides a new theoretical framework and technical methods for the field of knowledge tracing but also offers strong support for personalized learning services in intelligent tutoring systems. By simulating the three stages of memory flow, MFcKT can more accurately track the evolution of students’ knowledge and provide them with personalized learning resource recommendations and instant feedback. Additionally, the study provides new directions for future exploration in educational data mining and artificial intelligence.
Other Valuable Information
The study also compared the effects of different fusion layers (similarity-based fusion and attention-based fusion), showing that the attention-based fusion layer has greater advantages in improving model performance. Furthermore, the research team explored the training and inference efficiency of the model, providing references for future model optimization.
The research on MFcKT brings new breakthroughs to the field of knowledge tracing, and its innovative memory flow design and efficient dual-channel structure offer new possibilities for the development of intelligent tutoring systems.