Heart-rate adjustment is essential for risk stratification using QT-based ECG biomarkers, but conventional adjustment formulas, such as the Bazett and Fridericia corrections, assume a memoryless QT–RR relationship and ignore temporal adaptation. We introduce a system-identification-inspired framework to model QT as the output of a nonlinear dynamical system driven by current and past RR intervals, producing personalized dynamic ECG biomarkers of repolarization adaptation. For evaluation, RR and QT intervals were extracted from Holter ECG recordings of 418 post-myocardial infarction (post-MI) and control participants who underwent mental and physical stress testing. An LSTM model was pretrained globally and fine-tuned per subject to predict beat-wise QT from RR history. Cross-evaluation over memory lengths of 1–120 s showed finite QT memory, with performance plateauing beyond 40–60 s (MAE, 12.8±3.4 ms; correlation, 0.82±0.10). Subject-wise trained models were exposed to identical 16-minute exercise-like RR-interval acceleration/deceleration patterns to generate QT responses with hysteresis patterns. These responses were used to train and test post-MI classification on 282 training and 122 hold-out subjects. The method outperformed conventional QT correction features, increasing classifier AUROC from 62.6% to 67.1%. Thus, dynamic QT–RR models can provide interpretable, personalized biomarkers of repolarization adaptation for cardiac risk assessment.