<p>Artificial intelligence systems increasingly mediate safety-sensitive decisions across social, educational and digital environments. While behavioral data is often used to infer risk or intent, prevailing system designs frequently transform transient actions into persistent evaluative artifacts, coupling behavior with long-term identity and amplifying irreversible harm. Existing approaches to ethical AI governance predominantly rely on post-hoc policies, guidelines or accountability frameworks, leaving the architectural sources of such harm largely unaddressed. This paper argues that ethical failure in AI systems is not primarily a matter of insufficient principles, but of architectural design choices that privilege persistence, accumulation and reuse of behavioral representations. We introduce a Core Safety Substrate as an intermediate architectural layer that enables behavioral risk signaling without constructing durable identity-linked profiles. The proposed substrate enforces ethical constraints through design mechanisms including temporally decaying risk states, non-identifying trace correlation and human-centered escalation logic that frames intervention as interpretive judgment rather than automated enforcement. By embedding ethical limits directly into system architecture, the Core Safety Substrate reframes ethical AI as a design problem rather than a compliance exercise. This approach preserves contextual awareness while preventing the formation of irreversible reputational trajectories. The paper contributes a design-oriented framework applicable to socio-technical systems where early behavioral signals may otherwise exert disproportionate long-term influence. We argue that, in safety-oriented AI systems, controlled forgetting is not a technical limitation but an ethical requirement.</p>