The preferences-as-memory (PAM) framework assumes that decisions (or valuation judgments) are made by retrieving relevant knowledge (attitudes, attributes, previous preferences, episodes, or events) from memory in order to determine the best (or a good) action. This process is functionally equivalent to constructing predictions about one’s experience of the consequences of different actions, that is, what Kahneman and collaborators (Daniel Kahneman & Snell, 1990; Schkade & Kahneman, 1998) call predicted utility. Rather than accessing the equivalent of stable, continuous, and infinitely resolvable utility and indifference curves, we suggest that people make an attempt to retrieve past reactions and associations to similar situations. It is in the nature of human memory that such retrieval attempts do not always generate a single and precise answer. Because retrieval depends upon prior encoding, memory representation, memory query, situational context, and prior attempts at retrieval, the results may vary.
See also “Constructing Preferences from Memory” by Elke Weber and Eric Johnson