Maybe I was doing it wrong. I was actually creating the bayes database into /root/.spamassassin/ folder. As I'm with amavis integrated, the right folder seems to be /var/lib/amavis/.spamassassin.
So I used the --dbpath option pointing to this folder in sa-learn, and it seemed to increase de database, as the bayes_toks file increased almost 2MB.
After this, when I went to ISPConfig Panel and added some spamfilter rules in black/whitelist. Then the size of the bayes_toks file in the amavis folder just went back to the size it was before I ran the sa-learn on them, as I could see by the modification time.
After all... what's the right way of learning spam with ISPConfig?