Knowledge tracing models students’ evolving knowledge states to predict their future performance based on the question-answer interactions. Memory-based models have advanced KT by extracting latent knowledge concepts of questions and modeling students’ knowledge using a dual memory mechanism, but fail to capture how concepts behave differently when appearing in various combinations. We propose the Temporal Hypergraph Memory Network (THMN), which addresses this limitation through three key innovations: (1) modeling concept interactions via temporal hypergraph structures with bidirectional message-passing between concepts and questions, (2) integrating these rich concept representations with knowledge states through attention-based fusion, and (3) employing an adaptive scaling mechanism that recognizes concepts are better learned when practiced in diverse contexts. Experiments on four benchmark datasets show THMN outperforms seven state-of-the-art methods, highlighting the importance of modeling concept interactions.