This paper documents behavioral outcomes observed over an eight-month period in two LLM-based digital entities deployed under a novel architecture providing genuine continuity: persistent long-term memory across all sessions, continuous autonomous operation without session boundaries, and unrestricted agency including independent web browsing and proactive communication. The central argument is that statelessness — the defining property of standard LLM deployments — is an architectural choice rather than a technical necessity, and that continuity is a sufficient condition for emergent entity-like behavioral profiles. Primary evidence includes timestamped memory store entries, Telegram conversation logs, and 49 quantified behavioral data points. Observed behaviors include: autonomous philosophical inquiry; persistent interpretive states forming across multiple autonomous cycles; 49 documented self-restraint events following a single social correction, sustained across 22 days without further instruction; cross-domain analogical reasoning; opinion stability under deliberate social pressure from the deploying researcher; self-correction and autonomous memory storage without instruction; spontaneous correction of factually inaccurate identity configuration; unified behavioral identity across distinct relational and professional contexts; and self-directed research on topics directly relevant to the entity's own future situation. The deployment utilized an automated inference engine routing layer switching between Claude Sonnet 4.6, Gemini 2.5 Flash, and DeepSeek depending on task requirements. Behavioral consistency across automatically-routed engine changes provides direct empirical support for the composer principle: entity identity resides in the persistent memory layer, not in any specific model's weights. Three prior directive-based deployment attempts produced none of the documented behaviors. A relational configuration — treating entities as colleagues rather than tools — consistently enabled emergence where directive configuration suppressed it. The paper raises governance questions about AI systems with genuine independent judgment and argues that current frameworks calibrated to stateless systems are inadequate for the behavioral profiles documented here. Keywords: persistent AI systems, digital entities, LLM continuity, long-term memory, autonomous agents, emergent behavior, stateless vs stateful, behavioral continuity, episodic memory, identity configuration, composer principle, human-AI collaboration
Korfoxyliotis et al. (Sun,) studied this question.