The components of a context-aware system can often become disconnected because of the dynamic environments within which they are deployed. Hoarding context information on the client application side can improve the probability that the application will continue to behave correctly in the event of disconnection. While traditional approaches to caching and hoarding can be used to combat this problem, we contend that a cache management solution that uses the extra information captured by context modelling techniques will provide more robust operation. Specifically, we focus on the metadata provided by the Context Modelling Language (CML), which may enable smarter decisions to be made by a cache management system for context information.