Disentangled Representations for Causal Cognition