Windows and Linux don't leave unused memory around - both use it heavily for page cache. However what we have a problem here with is active/resident memory use and/or VM subsystem making no so smart decisions(this part seems to have gotten better with ML).

So the end result is that when Windows and Linux can be quite usable with 2GB RAM with a rotational disk, for similar workload OS X slows down quite a bit.