practical memory limits

An old rule of thumb was that you should have 6 times as much memory as
your dataset will take.  But I think pretty much everything has been
improved since then, so you should be able to get by with less (others
may be able to give a better rule of thumb these days).

You might want to look at the biglm package, it allows you to do
regression models with only a portion of your data loaded at a time,
allowing for pretty much any size of data set in a limited memory
situation.

Hope this helps,