When I read about cognitive models that use the Semantic Pointer Architecture (SPA), such as a model for general intelligence, I often see a component called the "clean-up memory"? What's the function of this component and how does it function vary from SPA model to SPA model?
SPA is used (among other things) for combining (binding) and extracting (unbinding) knowledge representations for processing. This is a (purposely) lossy compression. In the "Learning Rule Generation for Induction" case, the clean-up memory is used to convert a general transformation that is being learned (lots of different transformations convolved together) into a transformation that is already known. So basically it's associating the in-progress transformation to a known transformation.
This is why the "clean-up memory" is also called the "associative memory" in other SPA models and in Nengo 2.0.