Knowledge Tracing (KT) models student learning from interaction histories. Recent work has begun applying large language models (LLMs) to KT through prompting, option weighting, and test-time reasoning, but the resulting AUC remains substantially below ID-based KT. We propose Language Model-based Knowledge Tracing (LKT), a PLM-agnostic supervised KT framework that reformulates each interaction as a natural-language sequence of concept text, question text, and a response token, allowing any encoder-based pre-trained language model (PLM), or any decoder LLM after LLM2Vec adaptation, to be fine-tuned end-to-end under a unified protocol. Across three datasets, LKT achieves the highest supervised AUC among Language Model-based KT methods on every dataset, outperforming LOKT and Thinking-KT (best 0.7105 and 0.7185 on DBE-KT22) and frontier LLM prompting baselines (AUC near 0.5 to 0.6). Among the LKT backbones, ModernBERT provides the most cross-dataset stable performance and is our recommended deployment default. Beyond benchmark performance, we analyze LKT's behavior with performance comparisons and explainable AI (XAI) techniques, including attention maps, LIME, and t-SNE visualization of the [MASK]-position embeddings, to characterize which textual cues drive predictions and how the learned representation encodes correctness. Our work contributes a PLM-agnostic and interpretable KT framework that future PLM advances can directly plug into.