AIngram-AR: An Integration of Autoresearch and AIngram

AIngram-AR is a fork of Karpathy's Autoresearch specifically designed to run effectively on consumer hardware. This project has been tested on laptops equipped with an RTX 4060 GPU and 8GB of VRAM, providing users a practical solution for experimenting with machine learning workflows without the need for high-end infrastructure.

Core Features

• Integration with AIngram: Seamlessly combines the capabilities of Autoresearch with the persistent memory features of AIngram, facilitating uninterrupted learning cycles across loops.
• User-Friendly Documentation: Comprehensive guides and documentation are available, covering everything from setup to advanced usage. Key resources include:
  • User Guide: A complete walkthrough from setup to experimentation results, including architecture and CLI reference.
  • Multi-Agent Modes: Options for running multiple agents in different concurrency modes (mock, solo, swarm) while utilizing a shared AIngram memory layer.
  • Agent Programming Guide: Instructions for orchestrating agents in the recall/edit/train/remember loop.

Differences from the Full Autoresearch

Unlike the full Autoresearch implementation, which continuously loops through its training process, AIngram-AR provides a streamlined approach for initial experimentation with a single run:

uv run train.py

This command executes one timed experiment for a duration of 5 minutes and compiles results, offering a straightforward way to analyze training performance without getting entangled in endless loops.

Requirements and Setup

AIngram-AR requires:

• Python 3.11+ for its integration,
• Proper GPU setup as described in the upstream Autoresearch documentation.

Follow the provided instructions to set up two virtual environments for optimal functionality:

• The primary virtual environment for AIngram and the memory CLI should leverage Python 3.11 or higher, while the submodule environment should use Python 3.10 as per upstream requirements.

Training on Consumer Hardware

For those utilizing consumer-grade GPUs such as the RTX 4060, AIngram-AR offers modified training parameters to accommodate limited memory resources. To avoid OOM (Out of Memory) errors, adjustments such as lowering DEVICE_BATCH_SIZE and MAX_SEQ_LEN are recommended to ensure smooth operation during training sessions.

Conclusion

AIngram-AR stands as an accessible and efficient solution for users looking to explore the integration of autoregressive models and persistent memory in machine learning experiments on consumer hardware. The combination of Autoresearch and AIngram provides a unique framework for advancing machine learning research and development.