The idea of maintaining a persistent markdown wiki instead of stuffing everything into a vector DB resonates with me. We built AFFiNE (an open-source knowledge base) and one of the biggest UX pain points was that users couldn't easily inspect or trust what the system "knew." Human-readable flat files solve the transparency problem elegantly. Curious how you handle conflicting updates when two sources say different things about the same entity?