Local-first screen memory for Windows. Continuous screen OCR, app-gated audio via virtual mic device ("Engram Mic"), and hotkey-activated dictation replacing existing cloude based service. All powered by Whisper + RuVector semantic search. No cloud. No hot mic. No Alexa. Your memory, your machine.

Engram is a Windows application that demonstrates a key capability for autonomous AI agents: persistent, searchable memory of work context. Rather than relying on cloud services, it continuously captures screen content via OCR and meeting audio via local speech recognition, storing everything on your machine with semantic search powered by vector embeddings. This addresses a real bottleneck in agentic workflows—agents need to remember what they've seen and heard to make informed decisions across tasks.

The system works quietly in the background, triggered by hotkeys for dictation and accessible via a local dashboard. Built in Rust with 11 specialized modules, it implements several practices relevant to agent design: a safety layer that redacts personal information before storage, a vector index for semantic retrieval rather than just keyword search, and a modular architecture that separates concerns like audio capture, transcription, storage, and search. The project is currently 75-90% complete, with remaining work focused on integrating the audio pipeline, hardening security, and completing the user interface.

For professionals exploring how agents can gain situational awareness, Engram illustrates one solution to the memory problem: instead of building agents that forget context between interactions, you can give them access to a searchable record of what they've observed. This becomes particularly valuable for agents that need to work across multiple applications or understand meeting context—they can query their memory to answer questions like 'what was discussed about the budget last week?' or 'what options did the user select in that dialog?'