Miniworld

Miniworld is a LambdaMOO-inspired multi-agent simulation where AI-powered characters live, remember, and evolve in persistent text-based worlds. Built in Godot 4.4 with local LLM integration via Ollama.

🧠 Watch AI Think and Act

Meet Eliza and Moss, two AI agents with distinct personalities:

• Eliza: A curious conversationalist who asks thoughtful questions and seeks genuine connections
• Moss: A contemplative observer who speaks rarely but profoundly, with a long-term perspective

These aren't scripted NPCs—they use local LLM models to make real decisions based on their observations, memories, and personality profiles.

🎭 Key Features

AI Agents with Memory

• Agents observe events, record memories, and use semantic search to recall relevant information
• Just-in-time prompt generation ensures AI always has the freshest context
• Private reasoning system: agents think internally while displaying observable behaviors
• Memory integrity monitoring keeps the simulation healthy

Self-Aware Agents

• Agents can view their own personality profiles with @my-profile
• Self-modification: agents can update their personalities with @set-profile
• In-game help system allows AI to discover their own capabilities
• Property-based configuration makes all settings runtime-editable

Classic MOO Architecture

• Composition over inheritance: objects gain capabilities through modular components
• Event-driven observation: see what others do, react in real-time
• Uniform command syntax for players and AI: command args | reason
• Everything persists to human-readable markdown files

Build Your World

• @dig to create new rooms
• @exit to connect locations
• @teleport to jump anywhere
• @save to persist the entire world to markdown vault
• Modify saved files externally and reload them

Transparent AI

• @impersonate <agent> to see exactly what an AI perceives
• View the full LLM prompt, memories, and available commands
• Debug why agents make specific decisions
• Memory status indicators show system health

🎮 How to Play

1. Type commands in the input box: look, say Hello!, go garden
2. Observe AI agents as they think, speak, and move autonomously
3. Build the world with @dig and @exit commands
4. Explore memories with note and recall for semantic search
5. Become an AI with @impersonate to see their perspective

🛠️ Technical Highlights

• Godot 4.4 engine with custom three-panel UI
• Ollama integration for local LLM inference (privacy-first, no cloud required)
• TextManager daemon for hot-reloadable message templates
• Dynamic memory budgeting scales agent memories based on available RAM
• Vector embeddings enable semantic search over notes and observations
• CommandMetadata registry provides auto-discovering help system
• Markdown vault persistence for git-friendly world storage

🔮 Philosophy

Composition over Inheritance - Capabilities via components, not class hierarchies Uniform Objects - Players and AI use identical systems Event-Driven - Observers react to world changes in real-time Transparent AI - Debug tools reveal exactly what agents see and think Just-in-Time Context - AI prompts built fresh with latest memories Persistent Worlds - Everything saved to readable markdown Self-Aware Agents - AI can introspect and modify their own configuration Discoverable Systems - Learn through in-game help, not external docs

🚀 Current Status

Fully playable with:

• ✅ 29 commands across 7 categories (social, movement, memory, building, admin, query, self-awareness)
• ✅ Two distinct AI agents (Eliza and Moss)
• ✅ Semantic memory search with vector embeddings
• ✅ World building and persistence
• ✅ Memory integrity monitoring
• ✅ Self-awareness commands for agents
• ✅ Property-based runtime configuration

📚 For Developers

• Open source codebase with extensive documentation
• Recursive CLAUDE.md pattern for hierarchical project context
• Clear separation of concerns: Core, Daemons, UI, Components
• Modular component system makes adding new behaviors straightforward
• Callback-based async prevents timing bugs with LLM queue
• Collaborative testing workflow designed for human-AI pair programming

🎯 Use Cases

• AI Research: Study emergent behaviors in multi-agent systems
• Interactive Fiction: Create living stories with AI characters
• Education: Teach AI concepts through transparent, observable agents
• Worldbuilding: Prototype narrative worlds with autonomous inhabitants
• Experimentation: Test different AI personalities and interaction patterns

🌟 What Makes Miniworld Different?

Unlike chatbots that reset after each conversation, Miniworld agents:

• Remember everything they observe and can search those memories semantically
• Live continuously in a shared world with other agents
• Self-modify their own personalities and discover their capabilities
• Show their work with transparent reasoning and debuggable decision-making
• Persist their state to human-readable markdown files

This is a living laboratory for multi-agent AI interaction, wrapped in the nostalgic format of classic text MUDs like LambdaMOO.