ContextOS
The graph-based memory layer that prevents Groundhog Day in AI.
The Problem
Every AI session starts from zero. You re-explain your project. You re-introduce your team. You re-describe constraints the AI forgot overnight. Standard RAG retrieves keywords but misses relationships — it doesn't know that Project A is blocking Project B, or that you made a decision three weeks ago that affects the code you're writing today. The context window can't hold six months of history, and even if it could, you wouldn't want to stuff all of it in every prompt.
The Build
ContextOS runs as a background service alongside your AI tools. A Chrome extension monitors Claude.ai and ChatGPT sessions, detects idle time, and ships conversation transcripts to a local FastAPI server. An extraction pipeline (built around OpenAI gpt-4o-mini) pulls out facts, projects, people, decisions, and relationships as typed graph triples — then stores them in KuzuDB (an embedded property graph) and LanceDB (vector embeddings). Four API endpoints do the work: /ingest queues transcripts, /recall runs hybrid graph + vector search, /context returns a curated injection block formatted for a system prompt, and /brief diffs yesterday's graph state against today's for a morning summary.
What Makes It Different
ContextOS captures relationships between concepts, not just keyword similarity — which is what lets it answer 'what's blocking Project A?' instead of just 'find things mentioning Project A.' It's fully local: both databases are embedded (no Docker, no cloud), all data stays on your machine, and the server binds to 127.0.0.1. It predates Kernel and is the standalone proof-of-concept for the memory architecture Kernel later extended. Built before persistent AI memory was a mainstream product category.