try to keep the general density of your model the same across the topology
unless you know you’re going to need more detail in certain areas, in which case wouldn’t it be better to have denser meshes there, so that you can add detail at lower levels of subdivision, thereby using your memory more efficiently?