We present a new algorithm ICGE for incremental learning of extended Mealy automata computing over abstract data types. Our approach extends and refines our previous research on congruence generator extension (CGE) as an algebraic approach to automaton learning. In the congruence generator approach, confluent terminating string rewriting systems (SRS) are used to represent hypothesis automata. We show how an approximating sequence R 0 , R 1 , ... of confluent terminating SRS can be directly and incrementally generated from observations about the loop structure of an unknown automaton A. Such an approximating sequence converges finitely if A is finite state, and converges in the limit if A is an infinite state automaton.