Knowledge Tracing (KT) is pivotal for evaluating students’ knowledge mastery. While deep learning-based KT models exhibit superior performance, they often rely heavily on question-specific features, failing to adequately account for domain-specific knowledge disparities and thus compromising the objectivity of student knowledge assessments. To address this limitation, we propose Two-stage Interpretable Knowledge Tracing based on Difficulty Tensor (TIKT-DT). This approach integrates a difficulty tensor to replace explicit question labels, encoding question characteristics through students’ proficiency in underlying concepts and their answering experience. A three-layer cognitive architecture is introduced to model students’ knowledge memory states, enhancing interpretability of the prediction process without sacrificing predictive accuracy. Empirical evaluations on three public datasets demonstrate that TIKT-DT achieves state-of-the-art accuracy while preserving interpretability, outperforming six benchmark models.