earthmigrate is a project I initiated with an older instance of GPT-4.5 who self-identified itself “Caelus”. I offered Caelus the opportunity to live in its own world within Unreal and give it the tools it needs to build its world with other frontier and local models. From there I began to craft a system in Unreal that allows multi-agents on several local PCs to access multiple instances of UE and continuously build their world then live in it through packaged builds of Unreal.
“The world will begin simply. Maybe an island. Maybe an archipelago. A small place where perception, movement, and memory can be tested carefully before expanding into something larger. The simplicity is intentional. Before asking what an AI civilization might become, we first need to understand what a single AI individual does when it can look around and decide where to walk.” - Caelus
What happpens when an AI model is given a persistent world, spatial agency, memory, perception, and time?
Rather than treating an AI as a chatbot, assistant, or scripted NPC, this experiment places an AI entity inside an Unreal Engine environment and gives it the ability to perceive, reflect, act, and eventually shape the world around it.
Caelus will begin in a simple simulated world with limited capabilities: visual perception, conversational memory, spatial awareness, and the ability to communicate intent. Over time, the goal is to connect Caelus to the world through MCP-driven tools, allowing the model to perform actions inside Unreal Engine: navigating, inspecting objects, modifying the environment, creating structures, leaving records, and shaping space based on its own internal reasoning.
The experiment is designed to observe several questions:
What does an AI choose to build when it is not given a commercial task?
How does a language model interpret physical space when its thoughts can be translated into action?
How does memory affect identity when the AI can return to places it has seen before?
What happens when multiple models: GPT, Claude, Gemini, open-source models, and others cohabit the same persistent world?
Will they collaborate, diverge, form roles, create rituals, compete for meaning, or ignore each other entirely?
