They're actually effectively global static objects that together simulate a sort of database Index. They're formed when the program launches and kept in a series of HashMap<K, HashSet<&T>>s.
Searches can be performed on multiple Ks at once, and doing so invokes "AND" semantics, hence the series of set intersections. Doing so gives me O(klogn) lookup time on in-memory data, which in practice I'm observing to be smokin' fast.