This project introduces "Titans," a novel neural network architecture capable of learning to memorize information at test time, addressing the limitations of fixed context windows in LLMs. It offers a multi-agent demonstration platform with a native desktop UI, enabling researchers and developers to interactively explore and visualize this adaptive memory capability. The platform facilitates understanding how different AI agents reason about memory problems, providing a unique benefit for advancing LLM inference.

How It Works

The core innovation lies in the "Titans" architecture, which integrates a neural long-term memory module into transformer models via three proposed mechanisms: Memory as Context (MAC), Memory as Gate (MAG), and Memory as Layer (MAL). Its key differentiator is test-time learning, where the memory module adaptively updates its parameters during inference based on a surprise metric, allowing it to retain information that contradicts current knowledge without additional training.

Quick Start & Requirements

• Primary install/run command: Clone the repository, install dependencies with pip install -r requirements.txt, configure API keys in .env (copy from .env.sample), and launch with python main.py. Windows users can use titans.bat or titans.exe for a bundled, dependency-free experience.
• Non-default prerequisites: Python 3.9+ is required. API keys for desired LLM providers (OpenAI, Anthropic, Mistral, Groq, Google Gemini, Cohere, Emergence) are necessary for agent functionality.
• Links: GitHub repository: https://github.com/ai-in-pm/Titans---Learning-to-Memorize-at-Test-Time. Research paper: https://arxiv.org/abs/2501.00663.

Highlighted Details

• Features seven specialized AI agents, each powered by a different LLM provider (GPT-4, Claude, Mistral, Groq, Gemini, Cohere, Emergence) and embodying a distinct role within the Titans architecture.
• Includes a native Tkinter desktop UI offering real-time telemetry, interactive charts, live visualization, and a collaborative insights view for synthesized cross-agent analysis.
• Supports side-by-side agent collaboration, allowing users to observe how various LLMs reason about the same memory problem.
• Provides a zero-friction setup, running with a single command and gracefully handling configurations with only a subset of API keys.

Maintenance & Community

The project is maintained by "ai-in-pm". Community interaction is facilitated through GitHub Issues for bug reporting and feature requests. No dedicated community channels like Discord or Slack are specified in the README.

Licensing & Compatibility

• License type: MIT License.
• Compatibility: The MIT license is permissive, allowing for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

Agent functionality is dependent on the availability and correct configuration of provider API keys; agents without keys will be marked as unavailable. The Google Gemini integration may display non-fatal deprecation warnings, though the application remains functional. The project is presented as a demonstration platform, suggesting it may be experimental in nature.