This paper addresses the adoption problem of everyday AI agents through the Raised Companion Agent framework. Recent agentic AI systems can browse, plan, operate interfaces, and execute multi-step workflows, but practical deployment still faces a separate question: why would ordinary users return to an agent every day when no urgent task is pending? The paper defines this unresolved problem as the capability-retention gap. The central claim is not that current AI agents lack capability, nor that emotional companion systems are automatically beneficial. Instead, the paper asks what design conditions are required for agentic AI to become a safe, repeatable, life-integrated daily interface. The paper develops the Raised Companion Agent as a relational-gamified AI agent that the user names, raises, shapes, and gradually personalizes, and that later reciprocates care through proactive life support. The motivating design image is drawn from virtual pets and companion systems, but the framework is not a nostalgic digital-pet proposal. Digital-pet mechanisms are treated as interaction-design signals: care, growth, ownership, continuity, and return motivation. These mechanisms are then translated into an AI-agent architecture that can perform useful tasks while preserving a durable emotional return loop. Four minimum conditions define the framework. A Raised Companion Agent must support user-shaped growth, preserve persistent companion identity, provide consent-gated proactive life support, and reward beneficial offline behavior rather than screen time. These conditions separate the framework from ordinary chatbots, mascot assistants, digital pets, productivity dashboards, and engagement-maximizing gamified apps. The novelty is not companionship, gamification, or agentic execution alone, but their conjunction under a safety-bounded offline reward structure. The evidential structure is deliberately conservative. The paper does not claim that a deployed product has already achieved the framework, nor that a human user trial has validated its effects. It is a conceptual framework and design-science contribution. Its purpose is to define the design category, specify the minimum conditions, translate psychological mechanisms into measurable design variables, and provide a staged evaluation protocol that can later support, weaken, or falsify the framework. The safety layer is central rather than supplementary. The same mechanisms that make a companion agent attractive—attachment, ownership, memory, growth, mirroring, and surprise—can also create dependency, social displacement, sycophancy, over-trust, privacy risk, and manipulative engagement. The paper therefore distinguishes identity mirroring from agreement mirroring. A safe companion may mirror the user’s routines, goals, preferences, tone, and developmental history, but it must not mirror harmful beliefs, unsafe plans, distorted self-concepts, or delusional interpretations. The companion should feel personal without surrendering its independent epistemic boundary. The central safety claim is the offline-reward inversion. Conventional engagement design asks how to keep users inside the product. The Raised Companion Agent framework reverses this target: the strongest reinforcement should point outward, toward beneficial offline action. Walking, sleeping, studying, budgeting, preparing documents, attending appointments, resting, and reconnecting with people should grow the companion more than compulsive chatting. The framework therefore treats successful retention as meaningful return, not raw screen time. The architectural layer separates companion identity from model execution. The companion’s name, growth history, memory policy, safety profile, consent rules, continuity state, and user-shaped development should persist outside any single model version, provider, prompt, or interface context. Language models and tools may change, but the raised companion should remain operationally continuous. This LLM-agnostic identity structure supports model routing, safety escalation, update rollback, and anti-lock-in portability. The same separation also yields the platform claim. Because companion identity is independent of any one model or character expression, the same governed identity substrate can support licensed characters, original companions, public-service mascots, or minimal non-character interfaces. The paper therefore frames the Raised Companion Agent as an IP-agnostic companion platform rather than a single-character product proposal. Licensed intellectual property may provide a powerful onboarding path when properly licensed, tested, and governed, but it is not required for the framework. The evaluation layer is organized around comparison rather than persuasion. A future study should compare standard tool AI, chat-based companion AI, full Raised Companion Agents, and ablation conditions such as a companion without offline re-anchoring. The framework is supported only if the full system improves meaningful retention and offline outcomes without worsening dependency, sycophancy, privacy, human-contact, or high-risk advice endpoints. Higher retention alone is explicitly insufficient. The governance layer converts companion-AI risks into release gates. The framework requires memory control, consent comprehension, data minimization, audit logs, high-risk domain deflection, child and vulnerable-user safeguards, subgroup-specific stop rules, and incident response. Public or welfare-oriented deployments require special caution because the users most likely to benefit may also be most vulnerable to dependency or over-trust. A positive average result cannot override subgroup harm. The conclusion is that the paper does not prove that Raised Companion Agents work in deployment. Instead, it establishes a design category, an architecture, a safety logic, and an evaluation standard for testing whether everyday AI adoption can be built around safe relational continuity rather than capability alone. The result is best described as a design-science framework for converting agent capability into meaningful daily return while keeping attachment pointed outward, toward the user’s offline life. Keywords: AI agents, Raised Companion Agent, companion AI, relational gamification, capability-retention gap, persistent identity, offline reward, safety inversion, bounded mirroring, sycophancy, privacy by design, human-AI interaction, public digital services, IP-agnostic platform, LLM-agnostic architecture.
Taekyung Lee (Fri,) studied this question.