I tried the one posted by firstname.lastname@example.org
which seems much smaller than wire.h with no success. I'm still working on my own but I'll try to modify the other one (since it works) to see if I can make it take less RAM.
But if RAM is the issue this probably won't help because I would like to drive 2 cascaded displays. Maybe I could strip down the matrix library to handle exactly what I need. Or it might be time for a different CPU.