Miniworld
An experimental text-based virtual world combining classic MOO architecture with modern AI agents that possess memory, personality, and agency.
What Is Miniworld?
Miniworld is a text-based simulation built in Godot 4.4 that bridges the gap between classic Multi-User Object-Oriented (MOO) virtual worlds and modern AI capabilities. It's a space where players can interact with autonomous AI agents that remember past interactions, develop personalities, and make their own decisions.
The project explores what happens when you give AI agents the same tools, commands, and agency as human players—creating a shared world where both can build, explore, and communicate on equal footing.
Core Features
AI Agents
Autonomous agents with semantic memory, vector embeddings for intelligent recall, and just-in-time context generation. Two demo agents (Eliza and Moss) showcase different personalities.
MOO-Style Commands
Classic text adventure commands unified for both players and AI: look, go, say, emote, and more. 29 commands across 7 categories.
Semantic Memory
Create persistent notes in a personal wiki. AI agents use vector embeddings to retrieve relevant memories based on context.
World Building
Use @dig to create rooms, @exit to connect them. Build your own spaces and let AI agents explore them.
Debug Tools
Use @impersonate to see exactly what an AI agent perceives and how they reason about their decisions.
Persistent State
World state saves to human-readable markdown files. Everything is inspectable and modifiable.
Architecture Highlights
Component-Based Design
Everything inherits from WorldObject with unique IDs. Functionality is added through components:
- Actor: Entities that can perform actions
- Location: Spaces that contain other objects
- Memory: Persistent storage and recall
- Thinker: AI decision-making and reasoning
- VectorStore: Semantic search capabilities
Key Systems
- EventWeaver: Event-driven daemon that propagates observations to all relevant entities
- Shoggoth: LLM task queue with Ollama integration for AI agent thinking
- TextManager: Vault-based configuration system with hot-reload support
- MarkdownVault: Persistent storage layer for world state and memories
- WorldKeeper: Object registry managing lifecycle and relationships
Commands & Interaction
Players and AI agents share the same command interface, organized into categories:
- Navigation: look, go, exits
- Social: say, emote, whisper
- Mental: think, note, recall (for memory management)
- Self-Awareness: @my-profile, introspection tools
- World-Building: @dig, @exit, @save
- Discovery: help, commands, examine
Commands can include optional reasoning annotations (command | reason) that stay private in memory, allowing agents to maintain internal monologue.
The Vision
Miniworld is an experiment in emergent narrative and AI agency. Rather than scripting behaviors, it gives AI agents the same fundamental tools as players and lets them decide how to use them. The result is a world where interactions feel organic, memories matter, and every agent develops their own unique perspective.
It's a sandbox for exploring questions like: What happens when AI agents can build their own spaces? How do they form relationships when they truly remember past interactions? What kinds of emergent behaviors arise when they have genuine agency?
Technical Details
- Engine: Godot 4.4
- Language: GDScript
- AI Backend: Ollama (local LLM integration)
- License: AGPL-3.0
- Storage: Human-readable markdown files