Brain–computer interfaces (BCIs) enable direct communication between the brain and external devices, with promising applications in assistive communication and neurorehabilitation. Among BCI modalities, non-invasive EEG is particularly appealing, but its widespread deployment is hindered by severe inter-subject variability, non-stationarity, and the need for extensive user-specific calibration. Conventional neural decoders such as EEGNet and HTNet offer high within-subject accuracy but struggle to generalize across unseen users and are prone to catastrophic forgetting during sequential adaptation. This study proposes the Meta-Adaptive Hilbert Framework (MAHF) —a novel decoding architecture that unifies spectral representation learning, meta-initialization, and continual adaptation to tackle these core challenges. MAHF integrates a trainable Hilbert transform layer that adaptively extracts log-spectral power features from EEG signals, a meta-learning algorithm that learns a parameter initialization suitable for rapid cross-subject adaptation, and scaling-shifting adaptation layers that fine-tune the model efficiently while preserving prior knowledge. We evaluate MAHF on the BCI Competition IV-2a motor imagery (MI) dataset using a leave-one-subject-out protocol under low-data adaptation settings. MAHF achieves an average accuracy of 77.25% ± 3.25%, outperforming EEGNet (55.70%) and MUPS-EEG (72.50%), along with a higher F1 score (0.7677) and ROC-AUC (0.8483). It further retains 79.03% accuracy on previously seen users’ post-adaptation, demonstrating strong resilience to catastrophic forgetting. These results positions MAHF as a calibration-efficient, user-independent neural decoder with state-of-the-art performance in cross-subject EEG classification. The code is available on