Autonomys Agents: A framework for building autonomous AI agents

Autonomys Agents is an EXPERIMENTAL framework for building AI agents. Currently, the framework supports agents that can interact with social networks and maintain permanent memory through the Autonomys Network. We are still in the EARLY STAGES OF DEVELOPMENT and are actively seeking feedback and contributions. We will be rapidly adding many more workflows and features.

Features

• 🤖 Autonomous social media engagement
• 🧠 Permanent agent memory storage via Autonomys Network
• 🔄 Built-in orchestration system
• 🐦 Twitter integration (with more platforms planned)
• 🎭 Customizable agent personalities
• 🛠️ Extensible tool system

Getting Started

1. Install dependencies: yarn install
   • Windows users will need to install Visual Studio C++ Redistributable. They can be found here: https://aka.ms/vs/17/release/vc_redist.x64.exe
2. Create your character config: yarn create-character <your-character-name>
3. Setup character config:
   • All character configs are stored in characters/{your-character-name}/config
   • Update .env with applicable environment variables
   • Update config.yaml with applicable configuration
   • Update {your-character-name}.yaml with applicable personality configuration (See Character System below).
4. Run your character:
   • For dev purposes in watch mode: yarn dev <your-character-name>
   • For production build and run: yarn start <your-character-name>
   • For interactive CLI interface: yarn cli <your-character-name>

Interactive CLI Interface

The framework includes an interactive terminal-based UI for managing and monitoring your AI agent. To start the interface:

yarn cli <your-character-name>

Features:

• Real-time character status and workflow monitoring
• Interactive command input with keyboard shortcuts
• Task scheduling and management
• Live output logging

The interface provides a user-friendly way to interact with your agent, monitor its activities, and manage scheduled tasks, all within a terminal environment.

Known Issues

• This feature is in the very early stages of development and will be rapidly evolving. See #242 for more information on planned features and known bugs.

Examples

The following examples demonstrate the use of the framework and are available:

• Twitter Agent
• Multi Personality

Character System

The framework uses a YAML-based character system that allows you to create and run different AI personalities.

Creating Characters

1. Character related files are stored in characters/{your-character-name}/

2. Create new characters by running the create-character.ts script:

   # Create a new character
   yarn create-character your_character

Character Configuration

Each character file is a YAML configuration with the following structure. For an example character personality configuration, see character.example.yaml and for example parameter configuration, see config.example.yaml.

Context Size Management

The orchestrator includes a message pruning system to manage the LLM's context window size. This is important because LLMs have a limited context window, and long conversations need to be summarized to stay within these limits while retaining important information.

The pruning system works through two main parameters:

• maxQueueSize (default: 50): The maximum number of messages to keep before triggering a summarization
• maxWindowSummary (default: 10): How many of the most recent messages to keep after summarization

Here's how the pruning process works:

1. When the number of messages exceeds maxQueueSize, the summarization is triggered
2. The system creates a summary of messages from index 1 to maxWindowSummary
3. After summarization, the new message queue will contain:
   • The original first message
   • The new summary message
   • All messages from index maxWindowSummary onwards

You can configure these parameters when creating the orchestrator:

const runner = await getOrchestratorRunner(character, {
  pruningParameters: {
    maxWindowSummary: 10, // Keep 10 most recent messages after summarization
    maxQueueSize: 50, // Trigger summarization when reaching 50 messages
  },
  // ... other configuration options
});

This ensures your agent can maintain long-running conversations while keeping the most relevant context within the LLM's context window limits.

Autonomys Network Integration

The framework uses the Autonomys Network for permanent storage of agent memory and interactions. This enables:

• Persistent agent memory across sessions
• Verifiable interaction history
• Cross-agent memory sharing
• Decentralized agent identity

To use this feature:

1. Configure your AUTO_DRIVE_API_KEY in .env (obtain from https://ai3.storage)
2. Enable Auto Drive uploading in config.yaml
3. Provide your Taurus EVM wallet details (PRIVATE_KEY) and Agent Memory Contract Address (CONTRACT_ADDRESS) in .env
4. Make sure your Taurus EVM wallet has funds. A faucet can be found at https://subspacefaucet.com/
5. Provide encryption password in .env (optional, leave empty to not encrypt the agent memories)

Resurrection

To resurrect memories from the Autonomys Network, run the following command:

Options

• -o, --output: (Optional) The directory where memories will be saved. Defaults to ./memories
• -n, --number: (Optional) Number of memories to fetch. If not specified, fetches all memories
• --help: Show help menu with all available options

Examples:

yarn resurrect your_character_name                                  # Fetch all memories to ./memories/
yarn resurrect your_character_name -n 1000                           # Fetch 1000 memories to ./memories/
yarn resurrect your_character_name -o ./memories/my-agent -n 1000    # Fetch 1000 memories to specified directory
yarn resurrect your_character_name --output ./custom/path            # Fetch all memories to custom directory
yarn resurrect --help                            # Show help menu

While memories are being fetched, they will be added to the vector database named experiences in the background, located in the <your_character_name> folder within the data directory.

Testing

To run tests:

yarn test

License

MIT