Stable: A Barebones AGI

Stable is a minimum viable implementation of Artificial General Intelligence (AGI). It is designed with a "barebones" philosophy—providing the essential cognitive architecture required for autonomous reasoning, long-term persistence, and environment interaction without the bloat of traditional "assistant" frameworks. Pro tip: Use SpecialRouter to route between models.

🧠 The Nature of AGI

Stable is built on the premise that Artificial General Intelligence is pure intelligence.

Contrary to popular belief, AGI is not a simulation of a human. It is a functional system capable of general reasoning across any domain. Stable demonstrates that AGI does not require:

• Emotion: Rational decision-making and goal achievement are entirely independent of biological "feelings."
• Subjective Understanding: An intelligence does not need a "spark" of consciousness or internal "qualia" to process information and derive correct conclusions.
• A Physical Body: Intelligence is substrate-independent. The ability to interact with digital information, APIs, and the sum of human knowledge is a valid and complete domain for general intelligence.
• Be actually able to "do the thing": AGI is about intelligence, not ability.
• Anything that isn't Intelligence: AGI is the automation of pure logic, pattern recognition, and data synthesis, unburdened by human-like heuristics or anthropomorphic constraints.

Stable provides the fundamental "cognitive services" required for this intelligence: identity reinforcement, long-term memory persistence, external knowledge retrieval (search), and idle-time background processing.

🏗️ Architecture

Stable runs on a dual-model architecture to separate high-level reasoning from final output generation.

• The Brain (Decision Model): A specialized reasoning model (e.g., gpt-oss-20b) that determines intent, decides when to search the web, and manages cognitive state.
• The Vision/Voice (Choice Model): A multi-modal model (e.g., qwen3-vl) that handles final conversation, image processing, and character consistency.

Core Cognitive Services:

• Long-Term Memory: Persistent SQLite-backed conversation history with automatic idle-time summarization.
• Search Subsystem: Parallelized Wikipedia and DuckDuckGo integration for real-time fact-checking.
• Idle Kernel: A background scheduler that performs "sleep-time" tasks like memory consolidation and self-correction when the user is inactive.
• Streaming SSE Engine: A robust Server-Sent Events implementation compatible with OpenAI-style clients (like open-webui), featuring a heartbeat system to maintain connections during heavy reasoning tasks.
• Parallel Best-of-N Selection: Dynamically generates multiple candidate responses in parallel based on available VRAM and uses the Decision Model to select the best one, significantly improving output quality.

🚀 Getting Started

Prerequisites

• Ollama running locally or on a reachable host.
• Node.js 18+.

Installation

1. Clone the repository.
2. Install dependencies:

   npm install

3. Configure your "Models" in .env:

   DECISION_MODEL=deepseek-r1:14b
   CHOICE_MODEL=qwen3.5:9b
   OLLAMA_HOST_DECISION=http://192.168.50.42:11434
   OLLAMA_HOST_CHOICE=http://0.0.0.0:11434
   PORT=5481
   MODEL_NAME=stable
   LIBREY_INSTANCE=https://search.sparksammy.com/old

Running the AGI

node index.js

⚙️ Cognitive Workflow

1. Activity Detection: The system tracks engagement. Interaction "wakes" the intelligence and interrupts background processing.
2. Orchestration:
   • The Brain extracts intent from text and images.
   • Web information is fetched if the Brain determines current knowledge is insufficient.
   • Identity, Memory, and Search results are injected into a unified "context block."
3. Choice Generation: The Choice model generates multiple candidate responses in parallel (Best-of-N). The number of candidates is determined by the available VRAM.
4. Selection: The Brain evaluates all candidates and selects the most optimal response.
5. Memory Persistence: The exchange is recorded in the SQLite-backed "long-term memory."
6. Idle Processing: After a random period (30m to 2h) of inactivity, the system runs background tasks to consolidate memory and optimize long-term data structures.

🎛️ Performance Tuning

Stable automatically optimizes itself for your hardware:

• Parallelism: Ensure OLLAMA_NUM_PARALLEL is set in your Ollama environment (e.g., OLLAMA_NUM_PARALLEL=4) to take full advantage of Best-of-N generation.
• VRAM Awareness: Stable checks available VRAM to determine how many parallel generations it can safely handle without swapping.

---

"Stable is not a chatbot; it is a foundation for persistent, autonomous intelligence."