Early detection of dementia, particularly during the preceding mild cognitive impairment (MCI) stage, remains a critical challenge due to its subtle behavioral symptoms. While recent advances have combined deep learning with simple behavioral assessments to promising results, most existing approaches lack full automation, real-time capability, and robustness across diverse environments. We introduce a real-time, fully automated, multi-center cognitive screening system based on the dual-task paradigm, powered by a gait-based graph convolutional neural network. Designed for online operation, our system performs continuous assessment during motor and cognitive tasks, while simultaneously collecting data for iterative offline refinement of the underlying classification model. To ensure real-world applicability, we gathered 660 samples from five locations, encompassing a range of subjects, physical environments and camera hardware. The resulting data heterogeneity is overcome through a robust selection of standardized modalities, applying augmentation techniques specific to the periodicity of gait, and employing transfer learning to adapt across devices. We achieved strong performance with a sensitivity of 0.9394 and specificity of 0.9065, demonstrating the system’s robustness and accuracy in multi-center settings.