High-resolution ocean general circulation models (OGCMs) increasingly resolve high-frequency motions, including near-inertial waves, but conventional offline band-pass diagnosis requires long, high-frequency three-dimensional velocity records and substantial storage and post-processing. Efficient online estimation is therefore needed to diagnose near-inertial currents directly during model integration. Existing two-step and three-step online filters overcome this limitation by combining velocity records from only a few time instants. However, their limited frequency selectivity may reduce the accuracy of near-inertial velocity estimates, particularly at lower latitudes. Here, we develop a temporally causal multivariate spatiotemporal artificial-intelligence framework for the online reconstruction of global-scale, multilayer near-inertial currents in OGCMs, using only current and historical model states during inference. Rather than assuming a universally applicable fixed model, this study aims to establish a global-scale pretrained framework that can be efficiently adapted to different ocean general circulation models using limited target-model data. To assess cross-OGCM applicability, the model is pretrained using CESM simulations and transferred to LICOM through limited-data fine-tuning. A CNN-LSTM model which learns a nonlinear, state-dependent mapping conditioned on local forcing and background ocean conditions is trained to estimate the zonal and meridional near-inertial velocity components. Model performance is evaluated against the offline band-pass reference and the existing two/three-step filters. Model-interpretability techniques are applied to identify the multivariate, temporal, and spatial information learned by the CNN-LSTM and to generate physically testable hypotheses regarding near-inertial current generation and evolution. Robustness is further examined across different latitudes, vertical levels, seasons, and extreme events. Streaming inference experiments will further assess runtime stability, memory requirements, and computational overhead during model integration. More broadly, this study can establish a new paradigm for high-frequency ocean diagnostics, shifting from storage-intensive and model-specific offline post-processing toward globally pretrained, interpretable, and transferable online emulation.