Well, the autolearn=no is pretty obvious from the lines you posted. You seem to have the defualt autolearn scores, which is something like 0.5 for ham and 10 for spam. Most of those messages hit between the .5 and 10, so they won't be autolearned either way.
One I saw got 0, and should have been autolearned as ham (at least by the standard cutoff values - I'm not sure it was *really* ham). However that one got autolearn=failed. The reason for failure isn't clear from the lines you posted, and I don't immediately see any problems with your config. Does the Bayes database really exist, and is it accessible by whichever user is scanning mail? The debug output from one of those 'failed' messages would probably tell more about the problem. Loren