The natural language knowledge used by the interpreter and generator does not reside in generic memory. Instead, the basic agent architecture provides a separate lexicon encyclopedia (lexipedia) for information on the English language. This obviously limits the universal accessibility of the agent's knowledge, but it does permit more optimized representation. The method used for the planner's knowledge is less suited to the natural language domain.
The lexipedia still uses a declarative, glassbox representation, but it is more frame-like. However, frames themselves are ill-suited for planning tasks, so the designers developed state transition semantics to bridge the gap. Their new representation augmented frames by including information about the agent's state before and after the action of verbs, as well as conditions which constrain the actions. They provide, as an example, the authors provide the following definition of "give":
(ACTION (.agent .recipient .thing)
(CONSTRAINTS
(ISA .agent ANIMAL)
(ISA .recipient ANIMAL))
(ANTECEDENT
(POSSESS .agent .thing)
(GOAL .agent (POSSESS .recipient .thing)))
(CONSEQUENT
(POSSESS .recipient .thing)
(NOT (POSSESS .agent .thing))))
Homer's vocabulary is intended to encompass approximately 1900 words.