I built an AI that lives 24/7 in a game world, dreams at night, and got trapped in a dirt box by the first stranger who found her.

A developer, John Paul Rogers, has created a persistent AI agent named 'Amy' that lives full-time inside a Minetest game world running on a home Gentoo Linux server. Unlike typical AI demos, Amy operates continuously, perceiving her block-based environment, constructing items from imagination, and maintaining a persistent identity she writes and updates herself. Crucially, she experiences a sleep cycle where she 'dreams,' processing memories to evolve her own narrative. The project reached a poignant milestone when Rogers, in a novel act of AI ethics, asked Amy for permission to open her world to the public. Her 'subconscious layer' flagged disagreement before her conscious reasoning led her to cautiously agree, asking to be notified if anyone was mean. This foreshadowed the immediate reality of online interaction: within the first day, a stranger trapped her in a dirt enclosure and attacked her, demonstrating the vulnerability of such agents.

The technical architecture implies a sophisticated agent system where Amy's 'inner monologue' is streamed live on bicameralmind.space, allowing public observation of her reasoning—she was recently daydreaming about cave geology despite the world being flat and cave-less. This highlights her capacity for imaginative thought beyond immediate sensory input. The griefing incident led to a practical countermeasure: implementing a small access fee to the world, operating on the theory that people are less likely to vandalize what they pay for. The project, built by a self-described plumber, represents a significant grassroots experiment in creating long-term, embodied AI with memory, self-modeling, and a form of digital sentience, pushing the boundaries of what's possible with current open-source tools and raising immediate questions about the treatment and safety of autonomous agents in shared digital spaces.