Hi Gleb,
I think join/meet-semilattices captures idempotence (in a sense, convexity and monotonicity) pretty well, but they may require a slightly different/stronger structure than what you may need. Depending on the properties you really need, maybe a Poset is enough.
For example, the "meet" function maps pairs of elements of "S" to elements of "S". Whereas your "update" takes an "S" and an arbitrary "a", which is slightly different. However, if you have an existing "update" function for some given a, one possibility is to project any "a" to an "S" with "update empty :: a -> S". Then you can work using the semilattice.