It is well known that two properties are very important for associative memory. One of them is the property that the firing rate of key patterns is constant. The other is the property that the similarity among them is low. It is also known that a two-layer random neural net has a function of pattern separation, when the firing rate of the second layer is controlled at a low value. Therefore, we can use this neural net for producing patterns which are favorable for associative memory. We propose a model of associative memory with a net having a function of pattern separation and investigate the ability of recall in this model. From the result of our investigation, it is shown that a high recalling ability is obtained in this model.