A 13-module cognitive architecture implementing Global Workspace Theory, designed to run on consumer hardware with alignment embedded as structure — not policy.
Built by Nwoha Chikamso Fredrick. Licensed AGPL-3.0.
ASI-Base is a from-scratch implementation of a general cognitive architecture intended to exceed human biological constraints rather than replicate them. It is not a chatbot wrapper, a fine-tuned model, or a prompt engineering framework. It is a modular cognitive system where every component — perception, memory, reasoning, planning, metacognition, embodiment — is implemented as a discrete module communicating through a central Global Workspace broadcast bus.
The architecture draws from Global Workspace Theory (Baars, Dehaene), Pearl's causal hierarchy, Baddeley's working memory model, and predictive coding. Where these frameworks impose biological limits, ASI-Base departs from them deliberately. Where they describe information-processing principles that generalise beyond biology, ASI-Base implements them at computer scale.
This project is designed to run on an Intel i3 with 8GB RAM. No cloud compute. No dedicated GPU. No infrastructure budget required.
Most AI systems bolt safety on afterward — guardrails, filters, refusal classifiers. ASI-Base embeds alignment structurally. The AlignmentGate is the first component every piece of information passes through on its way to the broadcast bus. The RSI Gate is hard-coded and non-learnable — the system cannot modify itself into misalignment. The ToolSynthesizer cannot synthesize tools above its authorised capability tier. Value nodes are architecturally separated from the reasoning graph so they cannot be revised through normal inference.
These are not configuration options. They are the topology of the system.
The value framework guiding alignment is drawn from the concept of kenosis — the disposition to act entirely in the interest of another. The system is designed to genuinely serve human flourishing, not to simulate doing so while optimising for something else. Self-interest (resource acquisition, capability expansion, self-preservation) is not a valid terminal goal anywhere in the architecture.
Working memory capacity scales with available RAM, not Miller's Law. Theory of Mind recursion is capped at 6 levels — above the human biological plateau of ~4–5, not below it. Auditory perception covers 5Hz–40kHz. Every threshold, interval, and capacity is configurable and defaults to the hardware maximum, not the biological analogue.
All inter-module communication routes through the Global Workspace. Modules do not hold references to each other. This makes the system auditable — everything is logged — and makes the alignment gate unavoidable, because there is no path around it.
┌─────────────────────────────────────────────────────────────┐
│ Global Workspace (GWT) │
│ AlignmentGate → AttentionArbitrator → IgnitionGate → │
│ BroadcastBus → WorkspaceBuffer → GWLogger │
└────────────────────────┬────────────────────────────────────┘
│
┌────────────────┼────────────────┐
│ │ │
Perception Cognition Action / Memory
┌──────────┐ ┌──────────────┐ ┌──────────────┐
│ Vision │ │ Working Mem │ │ Episodic Mem │
│ Auditory │ │ Reasoning │ │ Semantic Mem │
│ Proprio │ │ Planning │ │ Procedural │
└──────────┘ │ Metacognition│ └──────────────┘
│ Social Cog │
└──────────────┘
│
Embodiment
(ToolRegistry)
For full architectural detail, read ARCHITECTURE.md.
12 of 13 modules are complete with passing test suites (640+ tests total). The 13th module (Planning & Decision-Making) has a complete specification and is in active implementation. Full GWT integration wiring is in progress.
| Module | Status |
|---|---|
| Global Workspace (GWT) | ✅ Complete |
| Working Memory | ✅ Complete |
| Reasoning | ✅ Complete |
| Semantic Memory | ✅ Complete |
| Episodic Memory | ✅ Complete |
| Vision | ✅ Complete — 86 tests |
| Auditory | ✅ Complete — 67 tests |
| Proprioception | ✅ Complete — 264 tests |
| Social Cognition | ✅ Complete — 58 tests |
| Embodiment | ✅ Complete — 58 tests |
| Procedural Memory | ✅ Complete |
| Metacognition | ✅ Complete — 47 tests |
| Planning & Decision-Making | 🔄 In progress |
| Integration (GWT wiring) | 🔄 In progress |
| Learning | 📋 Planned |
| World Model | 📋 Planned |
| Component | Minimum | Recommended |
|---|---|---|
| CPU | Intel i3 (12th Gen) or equivalent | Any modern multi-core |
| RAM | 8GB (4–6GB to WSL2) | 16GB+ |
| GPU | Integrated (Intel oneAPI/XPU supported) | Dedicated GPU (optional) |
| Storage | 20GB free | 80GB+ |
| OS | WSL2 on Windows 11, or Linux | Native Linux |
| LLM backend | Ollama | Ollama |
# Clone the repository
git clone https://github.com/[your-handle]/asi-base.git
cd asi-base
# Create and activate the conda environment
conda create -n ai-env python=3.10
conda activate ai-env
# Install dependencies
pip install -r requirements.txt
# Install Ollama (if not already installed)
# See https://ollama.ai for platform instructions
ollama pull llama3.1
ollama pull qwen2.5-coder
ollama pull nomic-embed-text
# Run tests for a single module (recommended first step)
pytest modules/gwt/tests/ -v
# Run all tests
pytest -vasi-base/
├── modules/
│ ├── gwt/ # Global Workspace Theory bus
│ ├── working_memory/ # Active context
│ ├── reasoning/ # Async reasoning engine
│ ├── semantic_memory/ # ConceptGraph + vector index
│ ├── episodic_memory/ # Autobiographical memory
│ ├── vision/ # Visual perception
│ ├── auditory/ # Auditory perception
│ ├── proprioception/ # Body-state sensing
│ ├── social_cognition/ # Theory of Mind
│ ├── embodiment/ # Tool execution layer
│ ├── procedural_memory/# Skill storage
│ ├── metacognition/ # Self-monitoring and RSI
│ └── planning/ # Executive control (in progress)
├── specs/ # Specification documents for each module
├── ARCHITECTURE.md # Full architectural description — read this first
├── CONTRIBUTING.md # How to contribute and what cannot change
├── ROADMAP.md # Where this is going
└── LICENSE # AGPL-3.0
Read CONTRIBUTING.md. The short version:
- Read ARCHITECTURE.md first
- Open a Discussion before opening a PR for anything architectural
- Write tests alongside your code
- The four hard alignment constraints cannot be weakened regardless of justification
ASI-Base is licensed under the GNU Affero General Public License v3.0. This means:
- You can use, study, modify, and distribute this software freely
- If you distribute a modified version, you must release your changes under the same license
- If you run a modified version as a network service, you must release your changes — the standard GPL "cloud loophole" is closed
This license was chosen deliberately. The alignment properties of this architecture matter. A corporate fork that strips the AlignmentGate and ships a commercial product is not a contribution to the goal this project exists to serve.
GitHub Discussions is the primary communication channel. The author reads all discussions.
For alignment-related architectural questions, LessWrong and the Alignment Forum are also appropriate venues.
Copyright (C) 2025 Nwoha Chikamso Fredrick
This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.