We've got a simple functional language here - being very unscientific, and looking at the key data structures, for my current workload, straight pointers are about 5-10% of the allocation. In practice I do see a 10-25% memory usage increase going from 32bit to 64bit (Linux).
(Anecdotally -- I've read 20% is a rough rule of thumb for 64bit anyway).
There is also an increase in data structure size due to data alignment, but this is going to vary quite a bit -- depends on the structure, the compiler -- and if you're doing dynamic memory allocation you may even find it makes no difference at all -- as malloc may have been delivering an oversize allocation anyway.
So, it's certainly not a trivial increase... Especially as we're currently running on a 256mb slice (I sort of wish they had a supported 32bit option)... but it's not massive.
However, for us, 64 bit still has a lot of advantages. For example, you can do more expansive memory mapping and the like.
Not sure if that helps.
arc runs in MzScheme 352 (http://download.plt-scheme.org/mzscheme/v352.html)
the freeBSD binaries on this page are i386, if thats what the new server is using then it won't make a difference. The source is available as well, if pg compiled the source in 64 bit mode it would increase memory usage.
But using a 64 bit OS doesn't necessitate using a 64 bit address space.
If your compiler is unfriendly, your structure size can change significantly -- afaik, GCC is pretty clever about structure packing, but I'd imagine it varies a lot depending on the architecture.