The connection manifest writes connections to disk so that a node can remember, after restart, what connections it used to have, even if they happen to be down right now. This allows it to buffer messages for a connection that might come back up soon.
As far as I can see, the connection manifest's update method only gets called when a node connects or disconnects. So on a stable network where nodes have remained connected for more than 24 hours, the watch_expire process will drop nodes from the manifest, even though they are still active.