ScreenMind captures your screen when content changes, understands it with a local Gemma 4 model, and turns it into searchable, chat-able memory. Ask "what was that error 2 hours ago?" and get the answer from your history. Also transcribes meetings, records voice memos, and lets you build automations in plain English. No cloud, no accounts, no telemetry. Runs on a 4GB GPU. MIT licensed. pip install screenmind.

Hey everyone 👋 I built ScreenMind because I kept losing context — switching between tabs, closing a terminal, scrolling past a message — and then spending 10 minutes trying to find what I just saw. The idea is simple: capture my screen when something changes, have an AI understand what I'm looking at, and make all of it searchable. The hard part was making it run entirely on my own machine. I didn't want my screenshots going to anyone's server. It uses Gemma 4 via llama.cpp, so it runs on consumer GPUs (tested on a GTX 1650). The search is semantic — you describe what you're looking for in plain language and it finds it. A few things I'm proud of: -Chat with your screen history — ask follow-up questions and it pulls context from your actual day -Write agents in plain English markdown — "every 6 hours, summarize my coding activity" — and they just run -MCP server built in, so Claude and Cursor can query your screen history directly -Auto-redacts sensitive data (credit cards, API keys) before storing anything It's rough around the edges — setup could be smoother, and I've only tested on Windows so far. Would really love feedback on what to improve first. Happy to answer any questions about the tech!