Millimeter-wave radar enables non-contact heartbeat monitoring, yet accurately reconstructing electrocardiogram waveforms from radar signals remains a challenge. To address the issues of large parameter sizes and high computational costs in deep learning-based reconstruction methods, which hinder edge deployment, and the limited representation capacity and insufficient accuracy of pure hyperdimensional computing (HDC) approaches, this paper proposes a teacher-student hyperdimensional learning framework (TS-HDL) that achieves high-precision ECG waveform reconstruction while maintaining low inference costs. First, raw radar I/Q signals are preprocessed to obtain chest displacement signals. After adaptive respiration suppression to remove respiratory interference, maximal overlap discrete wavelet transform is employed to extract components containing the main heartbeat energy as model inputs. The teacher network, a lightweight 1D convolutional neural network (CNN), learns the nonlinear mapping between radar signals and ECG to generate a task-oriented high-dimensional projection matrix. The student network, based on HDC, encodes inputs into binary hypervectors using this matrix and performs pointwise electrocardiogram prediction and online personalized fine-tuning via associative memory and Hebbian updates. Experiments on a public dataset demonstrate that under an 80/20 split, the proposed method achieves an average Pearson correlation coefficient (PCC) of 0.807 and a normalized root mean square error (NRMSE) of 0.082. For cardiac event timing, the median absolute errors for R, Q, S, and T waves are 5, 35, 22 ms, and 10 ms, respectively. In leave-one-subject-out cross-validation, NRMSE remains stable between 0.08 and 0.11, with PCC consistently above 0.6. During inference, the student model has only 0.02×106 parameters and a per-frame latency of 2.64 ms, reducing the parameter count by approximately 36 times and increasing inference speed by about 3 times compared with a convolutional neural network and long short-term memory (CNN-LSTM) baseline. By combining the strong representation capability of deep learning with the lightweight inference advantage of HDC, TS-HDL significantly reduces computational overhead while ensuring high reconstruction accuracy and supports online personalized updates, offering a feasible solution for edge deployment of non-contact ECG monitoring.