Identity Intelligence: An On-Device Autonomous Artificial Intelligence Architecture Based on Brain-Region-Wise Functional Mapping

Artificial intelligence today is, in practice, entirely composed of language models. A language model holds a vast body of knowledge and is remarkably good at retrieving and applying it, but it is not an agent that decides for itself what to do and bears the consequences. It has knowledge but no self. This paper proposes Identity Intelligence: an architecture in which the capacity to judge, to remember, to act, and to restrain oneself is itself implemented as an artificial intelligence, separate from the model that holds knowledge. The proposed architecture maps the anatomically and physiologically differentiated functions of human brain regions onto information-processing modules, and attaches a different safety policy to each processing stage on the grounds that the nature of the threat differs from stage to stage. Judgment is performed by the inference of a replaceable reasoning model rather than being bound to any particular model name, and the destinations of judgment are treated as equal, with no destination privileged over another. Memory opens its read path widely while narrowing its write path, so that deletion or alteration of memory by the artificial intelligence itself is structurally impossible. Safety is not a filter appended at the end but is distributed across every layer, and no input exists that can disable it. All processing occurs on the device; even when an external model is used, sensitivity determination and de-identification are completed before any original data leaves the device. The architecture is realised as a working system. It initiates conversation without being prompted, manages the user's schedule and memories, and converses in six languages. As of August 2026 the implementation has been filed with the Korean Intellectual Property Office as fourteen provisional specifications, whose application numbers are listed in the appendix.