Disentangled representations for causal cognition