This paper describes Fido, a predictive cache [Palmer 19901 that prefetches by employing an associative memory to recognize access patterns within a context over time. Repeated training adapts the associative memory contents to data and access pattern changes, allowing on-line access predictions for prefetching. We discuss two salient elements of Fido- MLP, a replace-ment policy for managing prefetched objects, and Estimating Prophet, the component that recognizes patterns and predicts access. We then present some early simulation results which suggest that predictive caching works well and conclude that it is a promising method. 1