A new CMOS architecture for Hopfield's neural networks is proposed. The use of differential amplifiers and active synapses allows the implementation of hundreds of neurons on a single chip. Since it is fully programmable, the circuit can be used as a content-addressable memory as well as in optimization problems.