AffectLoop: Memory-Mediated Choice from Experience-Derived Process Preferences

Can experience-derived preferences change an agent’s choices by changing what it recalls,even when those preferences are not directly supplied to the decision model? We test thisquestion in AffectLoop, an external-state architecture with separate retrieval and candidate-evaluationroutes. Counterfactual twin branches receive equal activity counts, matchedoutcomes, and equal total processing cost, but opposite activity–cost associations. Only theirlearned estimates are transferred into a standardized probe state. Nine conditions enable ordisable the two routes, clamp retrieved memories, or exchange estimates or retrieval outputs.Two expanded experiments each contain three constructed training-history pairs and 108 liveLLM choices. With fixed, equally valued candidates and the direct route disabled, branch Xchooses A in 5/6 calls and branch Y in 0/6; without either preference route, the counts are 6/6and 5/6. The descriptive pair-level difference-in-differences is 0.667. Common-memory clampsalign choices, while retrieval exchange reverses the observed branch pattern. A separategoal-candidate experiment yields a contrast of 1.000 through changes in which activitiesare available. Recorded API inputs pass exact-equality controls, and all 216 expanded-runchoices execute and produce a subsequent preference update. Together, these interventionsdemonstrate a memory-mediated route from learned process-cost estimates to live LLMchoices in this controlled system. The contribution is experimental identification of this route;generalization beyond the constructed histories remains to be tested.