i've been building a persistent context layer that captures screen + audio continuously using screenpipe and syncs to a vector db.
how are you handling the relevance scoring? time-based decay alone doesn't work because some old context is critical (architectural decisions) and some recent context is noise. I've left it up to claude code to reason based on the timestamps, and is helpful for the most part, but far from perfect
working with continuous OCR capture across 3 monitors using screenpipe. at 1.2fps you get usable text extraction but use 600mb-2gb ram.
biggest issue is OCR can't distinguish directionality - ie. if someone messages you, or you type "let's cancel the meeting" the text is identical but the intent isn't
curious how you handle the volume - for heavy users with 50k+ emails, are you processing everything client-side or hitting the gmail api for each message?
how are you handling the relevance scoring? time-based decay alone doesn't work because some old context is critical (architectural decisions) and some recent context is noise. I've left it up to claude code to reason based on the timestamps, and is helpful for the most part, but far from perfect