In this paper，a RAM associative memory(RAM—AM)with the nonlinear mapping ability is proposed．Every input pattern is divided into a number of sub—patterns which can address RAM to train the RAM—AM，its storage matrix is encoded through correlative matrix method．As a heteroassociative memory，the AM can succeassy recall high—order nonlinear problems such as XOR and the parities problems．The analysis shows that its mapping ability depends on the partition for input patterns．Finally，this paper theoretically proves that the proposed AM has higher signal—to—noise ratio than those of Hopfield—like models and realizability under the existence of a large number of neurons．The experimental results confirm its feasibility．