i noticed that as well,.. but the servers clear out memory and virtual memory once you send a new job..
at least that's what i am observing here.
Somehow this is a swapfile issue.. any program that once consumed a lot of physical memory over a longer time, also gets a lot of paging (VM).
you can add these columns to the taskmanager to see what's going on...
in this case it's rhino having 2.7gb paged..
now if you quit such an application you will here your hdd go nuts for minutes (the new intel-ssds
will solve that problem) as windows is cleaning the fragments from the swapfile... which basically stalls your computer to a great extend.
one might not want mxcl to cause that automatically once the render or the merging process finishes, but a manually executable "purge memory" command would be a good thing to prepare mxcl for the next render while nodes are idle, otherwise rendering is delayed (as you noticed) at rendertime due to swap cleaning.