A predictive cache system learns data access patterns to build a graph of related data items. A main cache hosted by a cache server is accessed by a cache client having a local cache. A request to the cache server for one of the related data items generates a response including some or all of the related data items, which are stored in the local cache and can be served locally for subsequent requests. Predictive cache behavior coexists with legacy non-predictive cache operation. Load on the cache server and aggregate latency are considerably reduced. Graphs can be specific to a particular context. Locally cached content can be distributed and reused by other cache clients. The predictive cache system is suitable for large-scale web applications. Variations are described.