MOSO: A Privacy-First Local Adaptive AI Assistant Platform

MOSO is a privacy-first, local-first adaptive AI assistant platform that runs entirely on the device. It grows by learning user behavior and adapting to preferences, while fundamentally protecting user privacy through local inference and a multi-level memory engine. The project's core architecture includes cross-platform applications (based on Flutter), a multi-engine inference runtime (supporting llama.cpp, ONNX, etc.), and a layered memory engine, aiming to address the pain points of mainstream cloud AI services such as data privacy risks, network dependency, high subscription costs, and limited personalization.

With the rapid development of Large Language Models (LLMs), users' reliance on AI assistants has increased, but mainstream cloud AI services have fundamental issues: data privacy risks (data uploaded to the cloud), network dependency (unusable without internet), high subscription costs, and limited personalization. The MOSO project addresses these pain points, aiming to provide a local AI assistant that truly understands users and learns continuously while protecting privacy.

MOSO Core adopts a flexible multi-engine inference design, supporting multiple inference engines to adapt to different hardware platforms and scenarios:

• llama.cpp: CPU-optimized lightweight inference, suitable for resource-constrained devices
• ONNX Runtime: GPU/CPU hybrid inference, balancing performance
• CoreML: Native support for Apple devices, efficient inference
• MLX: Framework dedicated to Apple Silicon, leveraging M-series chips for acceleration
• ExecuTorch: PyTorch mobile deployment solution, supporting model quantization optimization This architecture can automatically select the optimal inference scheme to ensure a smooth cross-platform experience.

MOSO's memory engine is a key feature, adopting a four-layer architecture:

1. Episodic Memory: Stores conversation history and events, recalling interaction content
2. Semantic Memory: Extracts conceptual knowledge, understanding the user's knowledge background
3. Procedural Memory: Records workflows and preferred operations, learning user habits
4. Preference Learning: Optimizes preference understanding through continuous interaction, achieving personalization This system is implemented using vector databases and RAG technology, providing a continuous conversation experience while protecting privacy.

MOSO takes multiple measures for privacy and security:

• Local Inference: All model inference is completed on the device, no network required
• Data Isolation: User data is stored in a local encrypted database
• No Cloud Dependency: Works normally even without an internet environment
• Optional Cloud Sync: Encrypted sync only when authorized by the user In addition, the project uses a source code viewable license, with transparent code for community review, balancing transparency and commercial potential.

MOSO provides an ideal solution for the following users:

• Privacy-sensitive users (lawyers, doctors, journalists): No risk of data leakage
• Offline workers: Full functionality available even without a network
• Long-term learners: Accumulates knowledge graphs, serving as a personal knowledge management assistant
• Developers/tech enthusiasts: Open-source architecture allows custom models and extended functions The project proves that local AI assistants can provide high-quality experiences while protecting privacy, offering a reference for similar projects.

Currently, MOSO has established a complete code repository structure (application layer, core runtime, memory engine, etc.) with a modular design. Future development directions:

• Improve cross-platform support and optimize native experience
• Expand memory engine capabilities to support complex knowledge graph construction
• Add support for more pre-trained models to lower the entry barrier
• Establish a plugin ecosystem to allow third-party extended functions