I'm not sure I understand this question, as the canonical memcached server isn't a JVM. Did you mean virtual machine by chance?
I can say that "down" can materialize in different ways and the client is designed so they are all eventually handled the same way. In one case, where the TCP connection is reset, it will automatically try to reestablish the connection in background until the server comes back online. In another case, where it just "goes away" a heuristic on receiving a number of continuous timeouts for a given node will eventually force the client to reset that connection, and then handle it approximately the same way.