As with Evans' program and Marrian vision, rule-based systems need to make different pieces of information explicit at different times.
When printing rules for their user, what's important is a textual representation that looks like natural language. But an inference engine can't work with this. The internal form of the rules must (for example) allow the inference engine quickly to access the conclusion and each part of the condition. So the system has two options. Either store every rule together with appropriate natural language text, or back-translate the rule to text on output.
Last week, I said that indexing is (in Schank's view) one of the main
issues in AI. An example here: in the backward chaining examples, the
system had to search the entire knowledge base for rules whose
conclusions were pertinent to a given question. E.g. to find, given
Resort, all rules with
Resort in their conclusion.
Efficiency can be raised by making the connection explicit, and building
some kind of index which leads you straight from each variable to all
the rules which mention it.