Knowledge Tracing Based on Memory Mechanism

Knowledge tracing aims to infer students’ mastery states of knowledge concepts through their interactive behaviors during learning, enabling personalized education systems to adaptively recommend materials. While existing models have improved predictive performance, their designs overlook core principles of human memory mechanisms. Learning and memory are intertwined cognitive processes: learning involves knowledge acquisition and integration, while memory controls information storage and retrieval. To address this gap, this study proposed a Knowledge Tracing Based on Memory Mechanism (MMKT) that analyzes interaction data by integrating the forgetting curve theory and spaced repetition theory. The model identifies three critical factors influencing memory retention strength: repeated time gap (time between consecutive interactions on the same concept), sequence time gap (time elapsed since the last interaction), and past trial counts (cumulative historical attempts). These factors are embedded into a dynamic key-value memory network, where a static key matrix encodes knowledge concepts and a dynamic value matrix updates mastery levels. An erase gate and an add gate are designed to control memory decay and consolidation, respectively. Experiments on three public datasets demonstrate that MMKT outperforms baseline models in accuracy (ACC) and area under the curve (AUC), demonstrating the effectiveness of memory mechanism modeling for enhancing knowledge tracing model performance.