One interesting approach is prevalence.
http://www.prevayler.org/wiki.jsp

This appears like it wouldn't work for large amounts of data. But for small data systems that need to be fast this seems like a great approach. It keeps all the objects in ram but it requires a particular programming style of serializable objects. It takes a snapshot of all the objects each night, and it saves a log file of the transactions throughout the day. This is why all the objects must be serializable. So as I understand it:

• all of your data for the life of your application must fit in ram.

I suppose this could be useful for activities. The data could be loaded from disk then logged as it was running then saved again when the activity shut down. And the log file would be a full transaction of everything important that happened.