What artificial minds would need to remember

Abstract

The debate over machine consciousness has been dominated by top-down approaches that extract indicator properties from theories of consciousness. Developmental cognitive psychology, which has characterized how self-related capacities are assembled in humans, has been largely absent. We propose the Autobiographical Architecture Model (AAM), a five-layer developmental framework in which embodied affective self-regulation, self-recognition, autobiographical memory, the working self, and narrative self-coherence are linked by asymmetric scaffolding. The first four, in their autonoetic components, are proposed as jointly necessary for autonoetic consciousness in Tulving’s sense, and the fifth is its elaboration into narrative identity. Current large language models show uncontrolled traces at Layers 2 and 4, fluent simulation at Layer 5, and no signature at Layers 1 and 3. AAM is an adequacy constraint on the functional roles a candidate architecture must realize, not a homology thesis. On the evidence to date, the gaps at Layers 1 and 3 are architectural, the gap at Layer 4’s source monitoring is one of training objective and readout, and neither is a matter of scale alone. The sharpest prediction concerns Layer 4. Current systems locate the passage that supports a claim when their context contains one, and AAM predicts that no four-way classification of how a representation arose is trained in or used by default. We state the dependencies component by component, derive predictions for developmental amnesia, autism, and memory ownership, position AAM against indicator-property frameworks and biological naturalism, and specify seven benchmark protocols with controls and a stated falsifier.