Knowledge tracing (KT) is a fundamental task in intelligent tutoring systems, aimed at assessing students’ knowledge states and predicting their future performance. However, the sparsity of educational data poses significant challenges for existing KT models, particularly in 1) question embedding learning, and 2) knowledge state modeling. To mitigate this issue, several KT models have incorporated contrastive learning (CL). However, traditional CL methods rely on random data perturbations to generate contrastive views, which may introduce noise and hinder the effective capture of intrinsic relationships within the data. To address this, we propose a novel Multi-lEvel joinT contrastivE learning framework for KT (METE). Our approach introduces a CL method that optimizes both question embeddings and knowledge states by generating contrastive views through structural transformations or by leveraging outputs from different neural networks applied to the same data. This strategy preserves the intrinsic information while extracting rich self-supervised signals from multiple contrastive views, effectively avoiding noise. We apply this method to both question embedding learning and knowledge state modeling. Extensive experiments demonstrate that METE consistently outperforms state-of-the-art baseline models.