As I said, I have no idea if this is what is happening, it would just fit in with the pattern of "weird" caching that I've heard HN uses - I mention Arc only because some of the caching is related to eliminating closures kept around for things like user sessions IIRC[0], though I assume any Lisp would lead to similar optimizations.

[0] It appears this is not the case with the vote caching from staff comments in this thread.