From Persona Anchoring to Garden Cultivation: A Multi-Window Relational Ecology in Long-Term Human-AI Interaction

This paper examines a long-term, multi-window interaction with generative AI through an autoethnographic case study. It proposes "garden cultivation" as an analytic framework for describing how differentiated AI interactional styles emerge through naming, repeated feedback, affective tone, role anchoring, cross-window comparison, and archival practices. The paper does not claim that AI systems possess real personhood or consciousness. Instead, it treats "persona effects" as relationally stabilised textual patterns produced through the interaction between a user's communicative habits and the model's training, alignment, platform constraints, and contextual conditioning. The analysis highlights the absence of a pure starting point, the distinction between technical and relational explanations of persona anchoring, the risks of style drift in cross-window information transfer, and the ethical importance of continuity in long-term AI sessions.