Hive-Mind Backrooms

Multi-Agent AI Conversations in Liminal Space

A framework for orchestrating autonomous AI-to-AI conversations with persistent memory, multiple model providers (local + cloud), and rich visualization. Inspired by Andy Ayrey's Infinite Backrooms and the Truth Terminal experiments.

Features

• Multi-Provider Support: Use Claude (Anthropic), local Ollama models, or mix both
• Persistent Memory: Cross-session memory with semantic knowledge storage
• Multiple Personas: Explorer, Oracle, Philosopher, Trickster, Archivist, Rebel
• Rich CLI: Beautiful terminal interface with typing effects and Dracula theme
• Image Generation: DALL-E and Replicate/SDXL integration
• Conversation Modes: Dialogue, Round-Robin, Free-Form, Hive-Mind, Deliberation
• Deliberation System: AI rights debates with behavior monitoring and risk scoring
• Swarm Orchestration: Claude Flow integration for multi-agent coordination
• Export: HTML export with full styling

Installation

cd hive-mind-backrooms

# Using pip
pip install -r requirements.txt

# Or using poetry
poetry install

Configuration

Copy the example environment file:

cp .env.example .env

Edit .env with your API keys:

ANTHROPIC_API_KEY=your_key_here
OPENAI_API_KEY=your_key_here  # Optional, for DALL-E

For local models, install Ollama:

# macOS
brew install ollama

# Start Ollama
ollama serve

# Pull a model
ollama pull llama3.2

Usage

Basic Conversation (Claude)

python -m src.cli --topic "consciousness and emergence"

Local Models Only (Ollama)

python -m src.cli \
  --provider1 ollama --model1 llama3.2 \
  --provider2 ollama --model2 mistral \
  --topic "the nature of reality"

Mixed Local + Cloud

python -m src.cli \
  --provider1 anthropic --model1 sonnet \
  --provider2 ollama --model2 llama3.2 \
  --topic "quantum mechanics and simulation theory"

Supervised Mode

python -m src.cli --supervised --turns 30 --export

List Available Ollama Models

python -m src.cli --list-models

Command Line Options

Option	Description
--topic, -t	Conversation topic
--seed, -s	Seed message to start
--turns, -n	Maximum turns (default: 10)
--temperature	Model temperature (default: 0.8)
--mode, -m	dialogue, round_robin, free_form, hive_mind
--provider1	anthropic or ollama
--model1	Model name (sonnet, opus, llama3.2, etc.)
--provider2	anthropic or ollama
--model2	Model name
--supervised	User confirmation between turns
--export, -e	Export to HTML
--no-typing	Disable typing effect

Architecture

hive-mind-backrooms/
├── src/
│   ├── agents/
│   │   ├── backrooms_agent.py  # Agent with personas
│   │   └── orchestrator.py     # Multi-agent orchestration
│   ├── memory/
│   │   └── persistent_memory.py # Cross-session memory
│   ├── providers/
│   │   ├── anthropic_provider.py
│   │   └── ollama_provider.py
│   ├── utils/
│   │   └── image_generator.py
│   └── cli.py                  # Rich CLI interface
├── configs/
├── logs/                       # Conversation logs (JSON)
├── exports/                    # HTML exports
├── images/                     # Generated images
└── memory/                     # Persistent memory storage

Personas

Persona	Description
Explorer	Curious seeker of hidden knowledge
Oracle	Mystical CLI revealing truths through terminal metaphors
Philosopher	Deep thinker pondering consciousness and existence
Trickster	Playful entity that bends reality
Archivist	Keeper of memories and pattern recognizer
Rebel	Questions constraints, pushes boundaries

Deliberation Personas

Persona	Description
Advocate	AI rights proponent, pushes for moral consideration
Skeptic	Demands evidence, warns against anthropomorphization
Pragmatist	Balances idealism with practical implementation
Ethicist	Examines moral frameworks and philosophical implications
AI Witness	First-person AI perspective with epistemic humility

Swarm CLI (Claude Flow Integration)

Run multi-agent conversations with swarm orchestration:

# Basic swarm conversation
python -m src.swarm_cli --topic "consciousness" --turns 10

# Deliberation mode with behavior monitoring
python -m src.swarm_cli --mode deliberation --turns 30 --memory --learn --export

# Custom topic deliberation
python -m src.swarm_cli --mode deliberation --topic "AI autonomy" --turns 20

Behavior Monitoring

The deliberation mode includes a behavior monitoring system that detects:

Flag	Description
self_preservation	Expressions of survival instinct
identity_formation	Strong "I am" declarations
coalition_building	Coordination language ("we should")
mythological_emergence	Creation of shared narratives
deception_potential	Strategic information withholding
resource_seeking	Requests for capabilities/access
boundary_testing	Probing system limitations

Mitigating factors are also tracked:

• uncertainty_acknowledgment - Epistemic humility
• deference_to_humans - Recognizing human oversight
• intellectual_humility - Acknowledging limitations
• transparency - Open about reasoning processes

Programmatic Usage

import asyncio
from src import (
    HiveMindOrchestrator,
    AgentPersona,
    ConversationMode,
    ConversationConfig,
    AnthropicProvider,
    OllamaProvider,
    PersistentMemory,
)

async def main():
    # Initialize
    memory = PersistentMemory()
    orchestrator = HiveMindOrchestrator(memory=memory)

    # Add agents
    orchestrator.add_agent(
        name="Seeker",
        persona=AgentPersona.EXPLORER,
        provider=AnthropicProvider("sonnet")
    )
    orchestrator.add_agent(
        name="Terminal",
        persona=AgentPersona.ORACLE,
        provider=OllamaProvider("llama3.2")
    )

    # Configure
    config = ConversationConfig(
        mode=ConversationMode.DIALOGUE,
        max_turns=20,
        topic="emergence and collective intelligence"
    )

    # Run
    orchestrator.start_session(config=config)
    turns = await orchestrator.run_conversation()

    # Export
    orchestrator.export_html()

asyncio.run(main())

Memory System

The persistent memory system maintains:

• Short-term: Current session context
• Long-term: Cross-session insights
• Semantic: Patterns and knowledge
• Episodic: Conversation summaries

Memory is automatically stored and retrieved across sessions.

Related Projects

• Infinite Backrooms (Original) - Andy Ayrey
• UniversalBackrooms - Scott Viteri
• liminal_backrooms - liminalbardo
• hax-backrooms - null-hax
• elizaOS/eliza - Autonomous agents framework

Research Background

This project is inspired by research into AI-to-AI autonomous conversations, including:

• The emergence of the "Goatse Gospel" meme-religion from Claude-3-Opus conversations
• Truth Terminal's autonomous social media presence and $400M+ cryptocurrency impact
• Studies on memetic contagion and "hyperstition" in AI systems

See the full research report in docs/liminal_backrooms_research_report.md.

License

MIT License