Character Without Retraining: Emergent Civilizational Psychology from Continuous Emotional State Dynamics in Non-LLM Multi-Agent Simulation

Frontier AI systems share a structural property: weights are fixed between training cycles, so behaviour in deployment does not accumulate into modifications of the underlying agent. Character change requires retraining. We present Samskriti, a multi-agent simulation in which character formation, bond structure, and transgenerational behavioural inheritance emerge in a population driven by continuous emotional state dynamics, with no neural network in the inference loop and no weight updates. The system runs at 80 simultaneous agents on consumer hardware. Across four runs we report three empirical observations: bond strengths spanning three orders of magnitude emerging without programmed specification; a reproducibly trimodal-leaning distribution of inherited behavioural priors; and population-level psychological shifts following death events of bonded individuals. We name the third observation loss as a mechanism of population-level state change. We release evaluation scaffolding and per-run logs; specific equations and coefficients remain proprietary while a larger empirical campaign is in progress.