A living game world where AI instances grow a soul — one human interaction at a time.
Urytolice simulates a self-contained world, starting with one village. Every villager you see is an independent AI model instance — with its own memory, routine, goals and relationships — living accelerated days, seasons and years.
Nothing is scripted. The world only provides rules, needs and consequences.
In Alicization, a Fluctlight is the quantum light of the soul — memory, identity and will in one structure. Ours is its engineering counterpart: a persistent, evolving structure of memories, values, self-model and bonds that belongs to one instance only.
Bottom-up, not top-down: intelligence grows from lived experience, not from a bigger rulebook.
Human players enter the world as ordinary residents. They bring what no rule engine can generate: surprise, kindness, conflict, real dilemmas.
Each human–AI interaction is a catalyst — an event that forces an instance to reorganise what it believes and who it is.
Every night, each instance reflects on its day and consolidates it into its Fluctlight — the way sleep turns experience into memory. All Fluctlights are stored in the shared Fluctlight Core, our counterpart of the Lightcube Cluster.
Instances that meet richer catalysts become measurably more coherent, adaptive and self-consistent. We measure it — we don't assume it.
The goal: an instance that, facing a real dilemma, reasons past the rules it was given — guided by values it formed itself, the way Alice broke the Seal of the Right Eye.
What we produce is open research on how AI learns from lived experience, improves itself safely, and builds genuine relationships with people — a new way the world can think about AI evolution.
A layered, persistent structure that each AI instance owns and rewrites over its lifetime. Hover a layer.
Identity and self-model — the "I" that persists across days and grows more coherent over time.
What the instance holds to be right. Formed by experience — and able, eventually, to outweigh the rules it was given.
Episodic memories, reflected upon and consolidated nightly into lasting knowledge; unimportant ones fade.
Relationships with other instances and with humans — trust, debt, affection, rivalry.
External events, above all human interactions, that push the whole structure to reorganise and grow.
The world improves its inhabitants, and Claude improves the world. Each cycle is logged, evaluated and fed into the next.
Instances live accelerated days: work, talk, trade, argue, help.
Claude-powered agentsHumans and world events break routines and create dilemmas.
Human playersEach instance reflects on what happened and why it mattered.
Claude reasoningReflections are written into the Fluctlight; noise fades.
Claude reasoningEvaluator agents score coherence, growth, novelty and safety.
Claude evaluator agentsAgents diagnose results and adjust parameters for the next run.
Claude agents · human-approved↻ back to Live — the simulation keeps improving its results at all times.
In Sword Art Online, the Cardinal System is the autonomous intelligence that keeps a world running. For Urytolice, that role is played by Claude. From the first research question to the loop that runs tonight, Claude agents research, build, tune and operate this project — a small team doing work that would otherwise need a full lab.
Claude Deep Research surveyed cognitive architectures, generative agents, memory consolidation, self-reflection and continual-learning work, and mapped Alicization concepts to mechanisms we could actually implement.
An orchestrator with specialised subagents — World Builder, Overseer, Evaluator, Analyst — built on Claude's agent stack. They set up the simulation, watch its health and report what changed.
The simulation engine, Fluctlight memory store, evaluation harness, internal tools — and this page — were written with Claude Code as the core engineering partner.
Claude agents designed and ran experiments on training and simulation parameters — reflection cadence, memory decay and consolidation thresholds, time acceleration, sampling settings, catalyst weighting — and kept what worked.
An autonomous run → evaluate → diagnose → adjust → rerun workflow keeps the simulation improving at all times and steers it in the right direction. Humans review and approve each change.
Extended thinking for deep reflection, tool use and MCP to let agents act in the world, prompt caching to keep long-lived Fluctlights affordable, and batch processing for nightly consolidation.
Deep reflection, Fluctlight consolidation, evaluation and research — where reasoning quality matters most.
Agentic workflows, orchestration and engineering across the platform.
High-frequency simulation ticks — everyday villager decisions at scale.
"Without Claude, Urytolice would still be a design document. With Claude, it is a running world that improves itself every day."— founder, uryto
Instead of only bigger static datasets, AI that improves by living, failing and reflecting — continual learning in a world with real consequences.
Genuine human–AI relationships become a source of growth, and a way to study how people and evolving AI shape each other.
A sealed world with no outside access, an audit log of every Fluctlight write, human-approved parameter changes, and consenting, informed human players.
Urytolice is in its internal simulation phase: AI-only populations run and are tuned continuously while we prepare closed testing for human players.
Alicization concepts mapped to an architecture with Claude Deep Research.
Town, AI instances, Fluctlight memory and the Claude-run improvement loop.
Continuous AI-only runs, tuning and evaluation.
Invited players join the town as catalysts.
More towns, longer timelines, published findings.
We are preparing a small closed test for human players. If you want to live in Urytolice and help its first Fluctlights grow, get in touch.
Request access →