Why do some functional languages need software transactional memory?

Question

Functional languages, by definition, should not maintain state variables. Why, then, do Haskell, Clojure, and others provide software transactional memory (STM) implementations? Is there a conflict between two approaches?