KNN "no learning involved": one probaby wants to cross-validate K at the least, if not learn the metric.

Some methods say online learning isn't applicable. As pointed out elsewhere, objectives for K-means and mixture models could be fitted with stochastic gradient descent. In general there is always an online option. For example, keep a restricted set of items and chuck out ones that seem less useful as others come in.

(Aside: I have a very introductory lecture to machine learning on the web: http://videolectures.net/bootcamp2010_murray_iml/ — not for anyone that knows the methods on this cheat sheat!)