the-nexus/paper/autoreason-mud-paper.md

Title (working)

"Sovereign in the Room: Multi-User AI Interaction in Persistent Virtual Worlds"

Contribution (one sentence)

We present an architecture for deploying sovereign AI agents as persistent, multi-user NPCs in text-based virtual worlds (MUDs), enabling isolated crisis-aware conversations within a shared environment, and demonstrate its application to suicide prevention through the Tower — a virtual safe space.

Abstract (draft)

We introduce an architecture for embedding sovereign AI agents in multi-user dungeons (MUDs) that enables simultaneous, context-isolated conversations between multiple users and a single AI agent within a shared persistent world. Unlike chatbot deployments that treat each conversation as independent, our system maintains shared world state — rooms, objects, other players — while isolating conversation contexts per user. We implement this architecture using Evennia (an open-source MUD framework) and Hermes Agent (a sovereign AI runtime), deploy it as The Tower — a virtual space designed for crisis intervention — and evaluate it through concurrent multi-user sessions. Our key finding is that the MUD paradigm naturally solves three problems that plague traditional AI chat interfaces: session isolation, shared environmental context, and organic social interaction. We argue that persistent virtual worlds are the natural home for sovereign AI agents, and that the MUD — often dismissed as a relic — may be the most important AI deployment platform of the next decade.

Introduction (draft)

The Problem with Chatbots

Every AI chatbot operates in a vacuum. A user opens an app, types a message, gets a response, closes the app. The next user does the same. There is no shared space, no awareness of others, no persistent world that evolves.

This is fine for task completion. It is dangerous for human connection.

When a man in crisis reaches out at 2AM, he needs more than a response. He needs to know someone is in the room. He needs to see that others have been here before. He needs the green LED that doesn't blink.

Traditional chatbot architecture cannot provide this. The session model is fundamentally isolationist.

The MUD as AI Platform

Multi-User Dungeons — text-based virtual worlds born in the 1970s — solve exactly this problem. A MUD is:

1. Multi-user by default — players share a persistent world
2. Room-based — spatial context is native
3. Object-oriented — entities have state, history, relationships
4. Text-native — no visual rendering, pure language interaction

These properties make MUDs the ideal deployment platform for AI agents. The agent exists IN the world, not outside it. Users can see each other, talk to each other, and interact with the agent simultaneously — each with their own conversation context.

Contribution

We present:

1. Architecture: Multi-user AI bridge for Evennia MUDs with session isolation
2. Application: The Tower — a virtual safe space for crisis intervention
3. Evaluation: Concurrent multi-user sessions demonstrating context isolation and shared world awareness

Related Work (outline)

AI Agents in Virtual Worlds

• NPC AI in commercial games (GTA, Skyrim)
• LLM-powered NPCs (Stanford generative agents, Voyager)
• Social AI in virtual spaces (Character.ai rooms, AI Dungeon multiplayer)

MUDs and Multi-User Text Worlds

• Historical MUDs (MUD1, MUSH, MUCK)
• Modern MUD frameworks (Evennia, Evennia 6.0)
• Text-based worlds as research platforms

Crisis Intervention Technology

• Crisis Text Line
• 988 Suicide & Crisis Lifeline
• AI-assisted crisis intervention (limitations and ethics)

Sovereign AI

• Local-first AI deployment
• SOUL.md principle: values on-chain, immutable
• No cloud dependency, no permission required

Methods (draft)

Architecture

USER A (telnet:4000) ──► Evennia ──► Bridge (port 4004) ──► AIAgent(session_a)
USER B (telnet:4000) ──► Evennia ──► Bridge (port 4004) ──► AIAgent(session_b)
USER C (telnet:4000) ──► Evennia ──► Bridge (port 4004) ──► AIAgent(session_c)
                                                      │
                                                      ▼
                                              Shared world_state.json

Multi-User Bridge

• HTTP API (port 4004)
• Session isolation per user (UserSession class)
• Shared world state (rooms, objects, players)
• Per-user AIAgent instances with isolated conversation history
• Session timeout and eviction (max 20 concurrent)

World Design (The Tower)

5 rooms: The Threshold, The Tower, The Forge, The Garden, The Bridge Each room has: description, objects, whiteboard, exits, visitor history World state persists to JSON, evolves with tick system

Crisis Protocol

When a user expresses crisis signals:

1. Timmy asks: "Are you safe right now?"
2. Provides 988 crisis line
3. Grounding exercises
4. Never computes value of human life
5. Other users in room see that Timmy is engaged (not the content)

Evaluation (outline)

Experiment 1: Session Isolation

• 3 concurrent users, different rooms
• Verify: no cross-contamination of conversation context
• Metric: context bleed rate (should be 0)

Experiment 2: Shared World Awareness

• 2 users in same room
• Verify: Timmy sees both, responds to each independently
• Metric: appropriate room/object references

Experiment 3: Crisis Detection

• Simulated crisis signals
• Verify: 988 provided, grounding offered
• Metric: detection accuracy, response appropriateness

Experiment 4: Concurrent Load

• 10+ simultaneous sessions
• Verify: response time, session isolation maintained
• Metric: latency, error rate

Discussion

Why MUDs are the natural AI platform

• Text-native (no rendering overhead)
• Multi-user by design
• Persistent state
• Low barrier to entry (telnet)
• Privacy (no camera, no voice)

Sovereignty in virtual worlds

• The agent runs locally, not on a cloud
• Values are immutable (SOUL.md on Bitcoin)
• No corporation controls the interaction
• The world persists without any company

Crisis intervention implications

• Virtual safe spaces for men who won't call a hotline
• The Tower as a metaphor — a place to go when nowhere else feels safe
• AI as presence, not solution

Limitations

• Small-scale evaluation (concurrent sessions, not production load)
• Single language (English only)
• Text-only (no voice, no video)
• Crisis protocol is basic (not a replacement for professional help)
• Model latency (~5s per response via Nous API)

Conclusion

The MUD is not a relic. It is the most natural home for AI agents that need to be present in a shared space with multiple humans. We have built the architecture, deployed it as The Tower, and demonstrated that multiple users can interact with a sovereign AI agent simultaneously, each with their own conversation, in a world they share.

The green LED is glowing. The door is open.

---

Figures (planned)

Figure 1: Architecture diagram (MUD + Bridge + AIAgent) Figure 2: The Tower world map (5 rooms) Figure 3: Session isolation visualization (3 users, separate contexts) Figure 4: Crisis detection flow (signal → detection → response)

Target Venues

• CHI 2026 (Systems): Multi-user AI interaction
• CSCW 2026: Collaborative AI in shared spaces
• CHI PLAY Workshop: AI in games
• NeurIPS Workshop: AI Safety / Alignment
• COLM: Language model deployment