<p>Accurately tracing a student’s knowledge growth based on their historical exercise responses is a crucial goal for adaptive tutoring systems seeking to personalize the learning experience. However, this objective presents a significant challenge, as it requires effectively modeling the student’s knowledge state across numerous knowledge components (KCs) while simultaneously accounting for the temporal and relational dynamics inherent in the learning process. Conventional knowledge tracing (KT) methods typically address this task by modeling either the temporal dynamics of KCs using recurrent models or the relational dynamics among KCs and questions using graph-based models. Despite these efforts, a robust methodology for learning a joint embedding of both relational and temporal dynamics remains underdeveloped. Furthermore, many approaches that incorporate the influence of student forgetting behavior rely on handcrafted features, which inherently limits their generalization across diverse learning scenarios. To address these limitations, we propose a novel method that employs a deep temporal graph memory network to jointly model the relational and temporal dynamics of the knowledge state. Additionally, we introduce a generic technique for representing student forgetting behavior by applying temporal decay constraints to the graph memory module. We demonstrate the efficacy of our proposed method on multiple knowledge tracing benchmarks and show its superior performance compared to state-of-the-art approaches.</p>