This avoids doing work that will never be used (since the resulting unread counts will never be sent in a /sync response). The negative of doing this is that unread counts will be incorrect when the feature is initially enabled.
/forget