Show HN: Running a vision model on every screenshot on-device (github.com)

ScreenMind has launched an open-source tool that processes every screenshot on a user's device using Gemma 4, a versatile AI model, to create a private, searchable "AI memory." Unlike Microsoft's Recall, which faced privacy concerns for its data handling, ScreenMind emphasizes local processing without cloud dependencies or telemetry, ensuring user data remains secure and private. Users can interact with their screen history through a conversational interface, enabling queries like retrieving messages from specific apps or getting summaries of meetings captured in real-time. The technical aspects of ScreenMind are impressive, featuring advanced capabilities like smart capture that detects screen changes, hybrid search combining semantic embeddings with keyword search, and a flexible analysis mode that allows users to prioritize speed or depth. Gemma 4 integrates vision, audio, and reasoning in one model, distinguishing it from competitors relying on multiple models. The architecture is optimized for local performance, facilitating fast responses and dynamic interactions without sacrificing security or user privacy, making it a significant advancement in the AI/ML space for personal data management.