In the real world, an agent is bound to occasionally encounter a situation that its designers and experiences have not prepared it for. What does a context-sensitive reasoner do in such cases? One alternative is to fall back on some default knowledge about how to set parameters, focus attention, select actions, and handle events. This is not a very appealing alternative, since it could potentially require a great deal of reasoning effort. Also, a context-sensitive reasoner will not (by design) carry as much information in the form of applicability conditions as will a ``normal'' reasoner.
A better solution results from having contextual schemas spanning a wide range of generality, organized in generalization-specialization hierarchies and retrievable by content. Even when a novel situation arises, the agent will find some c-schema, though the fit to the current situation may be less than perfect-and in the worst case, the c-schema found will be the top c-schema in the memory, which is essentially devoid of content. Or such a memory may able to provide c-schemas or cases representing analogous contexts. These can then be used in a manner similar to case-based reasoning. A second approach is to use perhaps the same memory scheme, but also have contextual schemas representing context such as ``not in a recognizable context''. These ``meta'' c-schemas could then guide the agent as it strives to cope with the novel situation. These ideas will be pursued in future work.
**Put in something about Medic and Orca??**