In terms of cache coherency, an actor’s state is only ever read in one thread at a time (which thread processes the actor’s mailbox), and usually this actor is going to process up to 30 messages at a time before transferring the thread to another actor. This is similar to how quants work in the regular Windows / Linux scheduler.
Therefore, you can view updates in the actors' memory as transactional - there is no way for two processors to simultaneously access active memory, since it is private and available only to one thread at a time. As a result of this, cache coherence is not a problem to start with, because the actor model forces the linearized read and write history to state.