Abstract Synaptic connection matrix encodes limited information. It is well known that neural network memory with storage prescriptions based on Hebb s rule will collapse as more patterns are stored. By requiring that old patterns are automatically forgotten and the memory recalls only the most recent ones, a new short-term neural network memory based on Y.Baram s fractal neural network is proposed. Comparison is made with Morris and Wong's method and the experimental results are shown to be rather satisfactory and encouraging.