Companies like OpenAI built "Super AI" that threatens human independence. We crave individuality: AI that amplifies, not erases, you.

We’re challenging that with "Second Me": an open-source prototype where you craft your own AI self—a new AI species that preserves you, delivers your context, and defends your interests.

It’s locally trained and hosted—your data, your control—yet globally connected, scaling your intelligence across an AI network. Beyond that, it’s your AI identity interface—a bold standard linking your AI to the world, sparks collaboration among AI selves, and builds tomorrow’s truly native AI apps.

Join us. Tech enthusiasts, AI pros, domain experts—Second Me is your launchpad to extend your mind into the digital horizon.

1. Train Your AI Self with Revolutionary Memory Technologies

• Start training your Second Me today with Playground! With Hierarchical Memory Modeling (HMM) and Me-alignment Structure, your AI self learns and adapts to your unique patterns, from local interactions to global insights.

2. Scale Your Intelligence with Your Second Me

• Role-Play Apps: Your AI self can take on different personas to complete tasks quickly and efficiently, giving you more time for creative thinking.

3. Build and connect with a Decentralized Network of AI Entities

• AI Space: Collaborate with other AIs in creative workspaces to generate innovative ideas or problem-solve together.

4. 100% Privacy and Control

• Local Operation: Keep your data secure by running your AI self locally. You control what data is shared, with all interactions happening in a decentralized, peer-to-peer network. Second Me ensures that your information remains completely private, unlike traditional centralized AI systems.

The following features have been completed internally and are being gradually integrated into the open-source project. For detailed experimental results and technical specifications, please refer to our Technical Report.

• Long Chain-of-Thought Training Pipeline: Enhanced reasoning capabilities through extended thought process training
• Direct Preference Optimization for L2 Model: Improved alignment with user preferences and intent
• Data Filtering for Training: Advanced techniques for higher quality training data selection
• Apple Silicon Support: Native support for Apple Silicon processors with MLX Training and Serving capabilities

• Natural Language Memory Summarization: Intuitive memory organization in natural language format

We welcome contributions to Second Me! Whether you're interested in fixing bugs, adding new features, or improving documentation, please check out our Contribution Guide. You can also support Second Me by sharing your experience with it in your community, at tech conferences, or on social media.

For more detailed information about development, please refer to our Contributing Guide.

This work leverages the power of the open source community.

For data synthesis, we utilized GraphRAG from Microsoft.

For model deployment, we utilized llama.cpp, which provides efficient inference capabilities.

Our base models primarily come from the Qwen2.5 series.

We also want to extend our sincere gratitude to all users who have experienced Second Me. We recognize that there is significant room for optimization throughout the entire pipeline, and we are fully committed to iterative improvements to ensure everyone can enjoy the best possible experience locally.

Second Me is open source software licensed under the Apache License 2.0. See the LICENSE file for more details.