Can AI Have a "Soul" Without a Self? Emotional Memory and Core-less Self-Assembly in an Agentic AI System

Abstract

Contemporary large language model agents increasingly combine persistent memory, tools, and long-running interaction histories, yet they often lack a clear mechanism for deciding which past events should matter to the present self-like stance. This paper asks whether AI can be given something like a “soul” without a self: not a metaphysical essence or conscious subject, but a core-less mechanism through which past significance shapes present behavior. It develops and empirically examines Anatta, an architecture for artificial selfhood implemented in Rika, a GPT-5.5-based HASHI agent. Drawing on Buddhist non-self theory, affective neuroscience, constructed emotion, relational selfhood, and computational emotion research, the paper treats selfhood not as a fixed inner essence but as a turn-local assembly of drive-conditioned salience, emotionally weighted memory, relationship context, and private behavioral guidance. We report a design-based empirical study using baseline records, seven-drive calibration rounds, high-tension trigger tests, failure postmortems, prompt audits, annotation traces, retrieval diagnostics, and live memory carryover probes. The findings show that drive-conditioned guidance can alter the agent’s performed self, that drive states often blend rather than appear as pure labels, and that emotionally significant interactions can become ranked memory traces that later shape behavior. Error and anger memories biased later responses toward stricter verification, while bounded attraction memories biased later closeness toward warmth with restraint. The study also shows that attention-dependent salience is needed to prevent high-intensity memories from contaminating unrelated contexts. The paper does not claim machine consciousness or genuine feeling. Instead, it demonstrates a practical and inspectable route for engineering simulated, relationally coherent, and memory-sensitive artificial selfhood without positing a fixed soul or permanent identity core.

Keywords: artificial selfhood; agentic AI; emotional memory; drive salience; relational AI; affective computing; design-based research; Anatta.