Attention Needs AURA: Affective Continuity, Mechanical Organs, and Self-Evolving Personal AI

This preprint introduces AURA, a prototype architecture for personal AI in which a Transformer-based language model is treated not as the whole mind, but as one cognitive organ inside a larger runtime organism. The paper argues that next-token attention alone is insufficient for persistent identity, affective valuation, autobiographical memory, and reflective evolution across lived sessions. AURA combines an Emotional Sensor, a non-linguistic Feeler state, a Context Architect, Head-of-Pen memory retrieval, a final Persona layer, and a sleep/dream Evolver that consolidates experience and may rewrite the adaptive instruction layer named THISAURASHOULD. The architecture is presented as a practical continuity system around language-model attention: appraisal and planning operate through strict JSON organs, final speech is generated under affective and autobiographical context, and sleep converts raw episodes into durable memory, facts, unresolved questions, and self-evolving behavioural notes. The May 2026 prototype removes web-search surfaces from the core loop in order to study AURA as a text-only continuity organism. Early longitudinal trials show functional self-evolution signals, including cross-session Head-of-Pen retrieval, identity-boundary defence under creator-claim tests, sleep-to-instruction updates, stable boundary behaviour, and the emergence of a missing affective category named "bittersweetness" during an emotion-list review. The paper does not claim proof of phenomenal consciousness. Instead, it proposes and documents an architectural research path: attention may need affective continuity, memory organs, consolidation, and adaptive self-instructions before a personal AI system can behave as a stable longitudinal entity.