This paper presents an auto-encoder architecture that transforms multi-agent observations into dynamic knowledge graphs, eliminating manual graph construction in heterogeneous multi-agent systems. We introduce: (1) an graph encoder mapping continuous agent observations to latent graph representations, (2) the k-NN algorithm with Gabriel pruning and adjusted density estimation for handling non-uniform density distributions, and (3) a composite loss function balancing reconstruction accuracy and graph sparsity. Experiments in PettingZoo’s simple_speaker_listener environment (100 runs) show our k-NN with Gabriel pruning achieves superior performance compared to Delaunay triangulation, beta-skeleton, and other baselines. The generated graphs capture meaningful agent relationships and form hypergraph structures suitable for multi-agent coordination, enabling scalable deployment in production ecosystems.