Soar's long term memory is represented uniformly as productions that cannot be directly examined by other Soar rules. Knowledge in long term memory is indirectly accessed when a pattern of working memory elements match the preconditions of a production. Thus a Soar agent is not able to directly manipulate the contents of the knowledge in its long term memory.
The black box representation allows for flexibility in that there is no architecturally fixed representation of knowledge structures. For instance, operators are not limited to the predefined add and delete list structure that is inherent in STRIPS-style planning architectures.
Go to a discussion of this property for multiple architectures.