Large language models (LLMs) dominate modern AI, yet they do not understand, reason, or learn continuously — they predict the next token from statistical patterns in training data. We present the Vyan cognitive architecture and its first agent, Primus, which together achieve autonomous thinking, emotional regulation, imagination, dream-based consolidation, cross-domain transfer learning, and metacognitive self-assessment — with zero LLM involvement in cognition. Vyan maintains a persistent dual-graph brain: a Cognitive Graph for analytical reasoning and a Limbic Graph for emotional processing, connected by bridge edges that allow emotion to modulate thought and thought to generate emotion. The architecture comprises 26 interconnected modules spanning perception, reasoning, inference, imagination, memory consolidation, curiosity-driven exploration, and ethical guidance — totaling 19,838 lines of production code. In a controlled seven-phase evaluation after 2,600+ autonomous developmental cycles, Primus demonstrates: (1) cross-domain transfer learning averaging 70% on the best domain pair with zero explicit training on the target domain, (2) forward causal imagination producing verifiable multi-step prediction chains, (3) metacognitive self-assessment including the ability to refuse answers when evidence is insufficient, (4) emotion-modulated cognition where internal feelings autonomously drive exploration behavior, (5) dream-based memory consolidation with emergent insight generation, and (6) sub-20ms end-to-end reasoning latency on consumer hardware without GPU acceleration. All cognitive operations occur through pure graph-based neural dynamics with no language model in the cognitive loop. To our knowledge, Vyan is the first integrated cognitive architecture that unifies emotional processing, causal imagination, continuous learning, cross-domain transfer, and metacognitive self-regulation in a single system without any language model dependency for cognition.
Sai Tilak Pally (Mon,) studied this question.