Magi Full
← Back to Magi
Responsible: Lake Watkins
Ecosystem relationships: SingularityNET Foundation, OpenCog Foundation, TrueAGI, EARTHwise Ventures
Status: Active. MAGUS milestones M2–M4 documented as complete in MeTTa. Several Magi Assistant components deployed; others code-complete or in development.
This card provides technical depth beyond the concise Magi index card. Magi is an applied AGI initiative developing neuro-symbolic frameworks on Hyperon, organized around four differentiators: transparency, consistency, plasticity, and corrigibility.
MAGUS Framework
The Modular Adaptive Goal and Utility System — goal-driven decision-making in MeTTa with the Overgoal mechanism, hierarchical goal architecture, Bach modulators, and ethical scenario validation (M2–M4 complete).
Overview
MAGUS (Modular Adaptive Goal and Utility System) is a modular framework for goal-driven decision-making in AGI systems, implemented in MeTTa on Hyperon. It combines the forward-thinking vision of classic motivational frameworks (Psi, MicroPsi, OpenPsi/MetaMo) with practical utility-AI patterns from commercial game development (Sims 4, Guild Wars 2), striking a balance between future-proofing for AGI adaptability and maintaining a straightforward human-in-the-loop interface.
MAGUS is designed for use within the PRIMUS cognitive architecture on Hyperon, supporting integration with ECAN, DAS, and the MeTTa language.
The Overgoal
The Overgoal is MAGUS's supreme evaluative mechanism. It continuously assesses all system goals using two criteria:
- Measurability: Goals must have clearly measurable satisfaction metrics
- Correlation: Goal fulfillment must positively correlate with satisfaction of other system goals
Goals failing these criteria are demoted; metrics that strongly correlate with cross-goal satisfaction can be promoted to goals in their own right. This creates a self-refining goal hierarchy where the system is motivated to evolve its own objective structure — analogous to human self-actualization, the Overgoal can never be completely satisfied, ensuring continuous growth.
Goal Architecture
MAGUS organizes goals in a hierarchical structure:
- Primary goals: Core objectives with measurable satisfaction metrics
- Subgoals: Decomposed steps toward primary goals
- Metagoals: Strategic value adjustments for coherence, exploration, and safety
- Anti-goals: Constraint enforcement with penalties, preventing undesirable behaviors
- Considerations and discouragements: Soft contextual factors influencing scoring
The framework uses Bach's 6-modulator framework (PAD emotion model + attentional modulators) to integrate affective state into decision-making, and a Scoring v2 pipeline that combines all components into unified action selection.
Milestones
| Milestone | Focus | Status |
|---|---|---|
| M2 | Goal fitness metrics — measurability framework (confidence × clarity), MIC correlations | Complete (19/19 tests) |
| M3 | Metagoals and integration — metagoals, anti-goals, Bach modulators, Overgoal, Scoring v2 | Complete (24/24 tests) |
| M4 | Ethical scenarios and research — scenario validation, ablation framework, AIRIS/HERMES integration patterns, reproducibility archive | Complete (5/5 tests) |
Design Heritage
MAGUS builds on Oliver Watkins's prior work on the Sophia robot's GHOST system and his patent "Data-driven goal modeling reevaluation for robots or virtual characters." The framework addresses a fundamental challenge identified by Asimov (rigid rule systems fail under pressure) and Goodhart (measures become poor targets once optimized for) by allowing goals to evolve through the Overgoal mechanism while maintaining alignment via anti-goals and human oversight.
Tools and Assistants
MCP-based AI assistant suite: GM Assistant (Claude + Foundry VTT), Discord bot, Foundry bridge, Smart Glasses interface, and Archive MCP server.
Magi develops a suite of AI assistant tools that serve as both practical products and testbeds for neuro-symbolic AI integration. The tools share a common MCP (Model Context Protocol) architecture, enabling AI agents to consume structured data from multiple sources.
Magi Assistant GM
An AI production stage manager for live tabletop RPG sessions, using Claude as the reasoning engine. The system consumes session data from three MCP servers (Discord bot, Foundry VTT bridge, Magi Archive wiki) and delivers real-time advice to the GM as whispered Foundry chat messages. Currently at v7 with 7 design iterations driven by live session post-mortems.
Key capabilities: scene look-ahead from session planning cards, phonetic matching for speech-to-text garbling (Soundex + Metaphone), GM hesitation detection, anti-echo policy, flowing-RP suppression, and proactive content delivery from beat cards. Deployed to production with 92 tests across 3 files.
Magi Assistant Discord
A Discord bot providing per-user audio capture, text channel monitoring, and real-time speech-to-text (Google Cloud Speech V2). Designed to produce clean, annotated datasets for training an AI Game Master. TypeScript/Node.js with SQLite storage, MCP server for AI agent access. Phase 1 (recording) deployed; Phase 2 (transcription + MCP) code-complete.
Magi Assistant Foundry
A Foundry VTT bridge sidecar that captures Fate Core game state (actors, chat, combat, scenes) from the GM's browser via WebSocket and exposes it through an MCP server. Supports scene monitoring, status effect integration, intercut checkpoints, and experimental token proximity reporting. Architecture: browser-side Foundry module → WebSocket → Node.js sidecar → MCP.
Magi Assistant Smart Glasses
A voice-controlled wearable remote interface for AI coding agents (Claude Code, Gemini, Codex) running on VITURE Beast XR glasses. Not a standalone LLM client — it is a hands-free remote to existing development terminals, with session switching, push-to-talk, and paged terminal display. Multi-repo architecture: coordinator (Node.js), Android app (Kotlin), shared protocol library (TypeScript with codegen). Connected via Tailscale VPN.
Magi Archive MCP
MCP server for the Magi Archive wiki (Decko platform at wiki.magi-agi.org). Provides AI agents with search, read, create, and update access to the wiki's card-based knowledge graph. Used by the GM Assistant for runtime wiki lookups and by Claude conversations for wiki maintenance.
Partnerships and Applications
EARTHwise ai-server joint venture (AIRIS, Tree AI, EAB), game development applications, and SingularityNET ecosystem integration.
EARTHwise Joint Venture
The ai-server is a shared technical surface connecting Magi and EARTHwise Ventures. Built on Python/FastAPI with PydanticAI, it integrates three AI systems:
- AIRIS (Autonomous Intelligent Reinforcement Inferred Symbolism): Game-playing AI agent for the Elowyn card game, using symbolic causal reasoning rather than statistical pattern matching
- Elowyn Tree AI: An onboarding chatbot designed for player retention in the Elowyn game, guiding new players through game mechanics
- EARTHwise Alignment Benchmark (EAB): A framework for AI alignment research through fine-tuning, supervision, and diagnostic benchmarking within the game environment
The server supports both local models (via Ollama) and remote API models, with a FastAPI web API and WebSocket chat server.
Game Development Applications
Magi applies AGI principles in game environments as constrained testbeds where alignment, cooperation, and strategic reasoning are measurable:
- Butterfly Galaxii: A tabletop RPG campaign used as the live testbed for the Magi Assistant GM, running on Foundry VTT with Fate Core rules
- Inkling: A game demo in development to showcase AI integration for investor and partnership discussions
- Elowyn: A tactical card battle game (via the ai-server partnership) serving as an AI alignment research testbed
SingularityNET Ecosystem Integration
Magi operates within the broader SingularityNET/ASI Alliance ecosystem. The MAGUS framework is implemented in MeTTa and designed for integration with Hyperon components (AtomSpace, PLN, ECAN). Strategic partners include the SingularityNET Foundation, OpenCog Foundation, and TrueAGI.
Note: The Neoterics brand (a metaverse-native population of early-stage AGIs originally planned as a Magi project) is owned by SingularityNET. Early documentation describes Neoterics as part of Magi's Phase 1; however, the brand and its development direction are under SingularityNET governance.
Status and Resources
Current implementation status, 6-phase strategic vision, four differentiators, implementation anchors, and primary sources.
Current Status
- Complete: MAGUS M2–M4 milestones (all tests passing), Magi Assistant Discord Phase 1 (deployed), Magi Assistant GM v1–v7 (deployed), Magi Archive MCP (deployed)
- Code complete: Magi Assistant Discord Phase 2 (transcription + MCP), Magi Assistant Foundry bridge
- In development: Magi Assistant Smart Glasses, Inkling game demo, ai-server AIRIS/Tree AI/EAB integration, Hyperon wiki ecosystem documentation
- Planned: Social agent platform (Phase 2), ownership marketplace (Phase 3), distribution platform (Phase 5)
Strategic Vision
Magi's 5-year plan progresses through six phases:
- SingularityNET Integration (Year 1, current) — MAGUS + Hyperon integration, initial validation
- Social Agents (Years 1–2) — Customizable AI companions for education, gaming, and environmental knowledge
- Ownership & Marketplace (Years 2–3) — Open-code/closed-data model with SingularityNET marketplace integration
- AI Game Master (Years 3–4) — Commercial AI GM products building on the Magi Assistant platform
- Distribution Platform (Years 4–5) — Personalized content discovery powered by neuro-symbolic recommendation
- Non-Gaming Applications (Year 5+) — Healthcare, urban planning, financial services requiring explainable AI
Four Differentiators
All Magi products are organized around four design principles:
- Transparency: Visible reasoning processes enabling understanding of AI decisions
- Consistency: Reliable, reproducible outputs building trust in critical applications
- Plasticity: Genuine ability to learn and adapt based on evidence and interaction
- Corrigibility: Meaningful correction capability enabling true human-AI partnership
Implementation Anchors
- magi-assistant-discord — TypeScript/Node.js Discord bot (recording + STT + MCP)
- magi-assistant-foundry — TypeScript/Node.js Foundry VTT bridge sidecar
- magi-assistant-gm — Claude-powered AI GM reasoning engine
- magi-assistant-smart-glasses — Kotlin/Android wearable app
- magi-assistant-coordinator — TypeScript/Node.js session orchestrator
- magi-assistant-common — Shared protocol types with schema-first codegen
- ai-server — Python/FastAPI backend (shared with EARTHwise)
- Magi Archive — Decko wiki at wiki.magi-agi.orgwith MCP server at mcp.magi-agi.org
Primary Sources
- MAGUS Core Framework Design Document (magi-archive: Neoterics+Magus)
- Magi: Applied AGI — 5-Year Strategic Business Plan (magi-archive: Neoterics+Metta+drive-docs+Magi-Applied-AGI)
- Neoterics: A metaverse-native population of early-stage AGIs. Medium / SingularityNET.
- Magi Weekly meeting transcripts (Dec 2025 – Apr 2026), available on Hyperon Wiki RawData.
Related cards: MetaMo Full · PRIMUS Full · MeTTa Full · EARTHwise
Tags
Discussion