This repository provides a locally runnable, low-cost implementation of Generative Agents, simulating human-like behavior for interactive scenarios. It targets researchers and developers interested in agent-based simulations and LLM applications, offering a cost-effective alternative to high-API-cost models.

How It Works

The project utilizes locally runnable, low-parameter language models to simulate agent behavior, deviating from the original paper's reliance on expensive, high-quality instruction models like GPT-3. This approach reduces operational costs and enables local execution, making the simulation more accessible. The current implementation simulates a Dungeons & Dragons town, offering a more flexible environment than the original paper's simpler scenario.

Quick Start & Requirements

• Notebooks can be run locally (Jupyter Lab recommended) or on Google Colab.
• Requires 16GB VRAM for optimal performance, with support for environments with <8GB VRAM.
• Notebooks are available in notebook/Release (stable) and notebook/WIP (latest features).
• A WIP library for game simulation is located at game_simulation.

Highlighted Details

• Focuses on low-cost, locally runnable models for agent simulation.
• Simulates a D&D town (Phandalin) for a more free-form environment.
• Aims to improve agent context compression in future iterations.

Maintenance & Community

No specific community links or notable contributors are mentioned in the README.

Licensing & Compatibility

The license is not specified in the README.

Limitations & Caveats

The current model's performance is expected to improve with the advent of next-generation instruction-tuned models. The README does not specify the license, which may impact commercial use or closed-source linking.