In the past week we've seen multiple instances completely unresponsive. When examining the instances further it shows that according to memory usage graphs they're reporting ~800MB of memory used.
However they aren't actually given that much heap to begin with. When restarted they will indicate their heap is 512MB, but the instance will soon climb to 800 once again, becoming slow and eventually no longer responding to requests.

Workaround
Create a heapdump of the application that's using too much memory.
Provide a higher heap than the instance is using.