What it does

Second Me is a local-first system for training a personalized AI on your own data. It uses Hierarchical Memory Modeling and a “Me-Alignment Algorithm” to fine-tune Qwen2.5 models into something meant to represent you—your context, your identity, your quirks. The resulting “AI self” runs locally via Docker and can optionally connect to a decentralized network where other people (or apps) can interact with it, with your permission.

The interesting bit

The project treats personal AI as identity infrastructure, not just a chatbot wrapper. The academic backing—two arXiv papers on “AI-native Memory”—suggests they’re trying to formalize how an LLM encodes a person, rather than just bolting RAG onto a generic assistant. Whether that actually produces a convincing “you” is, of course, the open question.

Key highlights

• Fully local training and inference; data never leaves your machine unless you opt into the network
• Docker-based deployment with a web UI at localhost:3000
• Supports MLX acceleration for Apple Silicon (CLI-only)
• Built on open-source stack: GraphRAG for data synthesis, llama.cpp for inference, Qwen2.5 base models
• Apache 2.0 licensed, ~15.6k GitHub stars

Caveats

• Hardware requirements are steep and lopsided: 8GB RAM gets you a ~0.8B parameter model (barely functional), while even 32GB tops out around 3.5B—fine for a phone assistant, not exactly a deep thinker
• The README itself notes “significant room for optimization throughout the entire pipeline”
• Mac users get noticeably worse performance per-GB than Windows/Linux; M-series MLX support is CLI-only
• “Cloud solutions” are explicitly on the roadmap for May 2025, acknowledging local hardware is a bottleneck

Verdict

Worth a weekend if you’re curious about personal model fine-tuning and don’t mind the hardware tax. Skip it if you expect ChatGPT-level capability out of a 1.5B model running on your laptop—the project is upfront that this is prototype-grade.