Sheaf-based Federated Representation Learning (SFRL) optimizes local objectives via a manifold-constrained geometric alignment regularizer using learnable sheaf restriction maps. This framework enables heterogeneous agents to achieve global consistency through orthogonal transformations and isometric embeddings without assuming a shared global latent space.