Bayesian Corpus size

[pootwaddle]

What is the largest size of the Bayesian corpus that you recommend?
I read that the larger the corpus, the more accurate. I am finding that to
be the case as my ham/spam messages approach 5,000.

I plan on continuing to capture both ham and spam up to x limit, then
deleting the oldest as I retain x newest messages.

Any thoughts or recommendations?

Thanks!