Toward Memory-Aided World Models: Benchmarking via Spatial Consistency