Initiating Conversation: A Proactive Persona Agent with Relationship-Tiered Memory and a Simulated Inner Life

Most conversational AI systems are strictly reactive, responding only when prompted and remain-ing dormant between user sessions. However, in human relationships, initiating contact serves as afundamental social signal of attention, shared memory, and care. In this paper, we present Liliana, anLLM-based persona agent deployed in a live Discord community that autonomously determines when,to whom, and about what to initiate conversation. Our approach relies on three coupled mechanisms:(1) a simulated inner life, implemented via a scheduler that maintains ongoing agent activities anddaily rituals (e.g., morning reports and nightly diaries) to provide contextual reasons for reachingout; (2) relationship-tiered memory, which tracks persistent per-user state (relationship tier, mood,and a topic ledger) to govern conversational tone, initiation frequency, and social boundaries; and(3) artifact-grounded proactivity, through which the agent creates small digital works (micro-gamesand seeded generative art), shares playable links, solicits user feedback, classifies responses, anddeploys code fixes in a closed loop. We report the system architecture and preliminary telemetryfrom a first-week in-situ deployment in a small private community (N = 4 recurring users). Earlyobservations show a 77% response rate within a 12-hour window (n = 13 messages, largely drivenby interactions with the developer) and an 80% topic-grounding rate among measurable messages(n = 5). We also document practical deployment failure modes, including cross-user mood con-tamination and text-to-speech hallucinations caused by decorative punctuation, alongside the ethicalguardrails required for proactive companion agents.