Without knowing anything about your application beyond what you've posted here (and I admit, I skimmed a bit), I would rather optimize the handling of multiple "very, very large log files" than worry about object representation. This sounds more like a working set problem to me.

If you can do line- or chunk-at-a-time preprocessing into a small database (I like SQLite), you can likely reduce your working set size and forget almost everything about needing to manage it. I've used that successfully in what I believe to be a similar situation.