Predicting movement in multi-agent continuous systems, such as football, presents significant challenges due to the dynamic and interactive nature of the environment. This work proposes a novel approach to movement prediction by leveraging a graph-unified representation, where football players are modeled as nodes and their interactions in time and space as edges. The proposed architecture, GuardiolAI, integrates Graph Neural Networks (GNNs) with Generative AI techniques, specifically Variational Autoencoders (VAEs), to capture both spatial and temporal dependencies in a unified manner. Unlike traditional methods that process spatial and temporal data separately, our approach models these aspects concurrently within a single graph structure. The methodology involves encoding tracking data from possession sequences into graph-based representations and employing GATv2 layers to learn adaptive attention weights across spatial and temporal dimensions. The model is evaluated using standard movement prediction metrics such as the average displacement error (ADE), mean squared error (MSE), and final displacement error (FDE), and is compared against a constant-velocity baseline. Experimental results demonstrate that the proposed approach achieves competitive performance and provides insight into player interactions and team dynamics.