Here's a possible organisation (page 296). MOP = memory organisation packet. Describes a concept (generalisation over several episodes), and contains two-level index. Level one is feature type; level two is feature value. Index points at sub-MOPs or individual episodes.
So reconstructive remembering entails creating sets of keys and following these down from a starting MOP. Needs strategies for generating plausible key-sets and starting MOPs.
Examples:
(1) ``Have you ever discussed SALT with Gromyko at a diplomatic meeting?''. Can go from MOP1 via topic or participant. (Paths need not be unique.)
(2) ``Have you ever attended a diplomatic meeting in the CDA, with Dayan?''. Can go from MOP1, via topic or participant. This one needs more than one level of retrieval.
A problem: question may not specify a starting MOP. E.g. ``Have you ever discussed the CDA with Dayan?''. Assume that ``talking to people'' is not a MOP. So there's no starting point. But if memory contains some auxilliary classificatory information (page 304), then you can go from the description of CDA back to the general topic of international contracts, and then to a MOP. Schank proposes a number of retrieval strategies for use where MOPs or features are not specified, or are not sufficient.
This work demonstrates that there's a computational mechanism that can perform in some ways like human memory, and it shows how learning and retrieval are interlinked. Thus, like so much AI, it suggests an approach: a possible mechanism which the psychologist can try to refine or disprove. It certainly does not prove memory is organised in this way, and it fails to model a number of properties. One is that people know what they know and don't know: ``Do you know Caesar's birthday?'', ``No, and I know there's no point in searching my memory for it.''.