Ax0n: Model-Agnostic Think & Memory Layer for LLMs

Ax0n is a model-agnostic Think & Memory layer for LLMs. It enables structured, parallel reasoning with real-world grounding and persistent memory—no MCP needed.

Features

• Structured Reasoning: Multi-step thought processes with JSON meta-control
• Parallel Execution: Tree of Thoughts / APR-style branching and merging
• Real-world Grounding: Fact verification with citations and evidence
• Persistent Memory: Mem0-inspired knowledge extraction and storage
• Model Agnostic: Works with any LLM (OpenAI, Anthropic, local models)

Quick Start

pip install axon

from axon import Axon

# Initialize with your preferred LLM
ax = Axon(llm_client="openai", api_key="your-key")

# Generate structured thoughts
result = await ax.think(
    "What's the best time to visit Kyoto?",
    max_depth=3,
    enable_grounding=True
)

print(result.answer)
print(result.trace)  # Full reasoning trace
print(result.citations)  # Evidence sources

Architecture

Ax0n consists of 7 core modules:

1. Retriever - Context fetching via embeddings and KV lookup
2. Think Layer - Structured, parallel thought generation
3. Grounding Module - Real-world fact validation
4. Memory Manager - Knowledge extraction and persistence
5. Renderer - Output formatting with traces and citations
6. Orchestrator - Module coordination and flow control
7. Testing & Validation - Comprehensive test suite

Documentation

• Installation Guide
• API Reference
• Architecture Overview
• Examples

Contributing

We welcome contributions! Please see our Contributing Guide for details.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Inspired by Mem0's memory extraction patterns
• Built on Tree of Thoughts and APR research
• Community feedback and testing