Some modeling algorithms (naive bayes and local models, like k-nearest neighbor or kernel regression) can be updated immediately. In some sense, they can be used for recall and training very close in time.