Cognee is already good
On Cognee’s HotPotQA head-to-head their graph scored 0.93 human-like and 0.84 F1. On BEAM at 100k tokens they report 79% against a Llama-4 RAG baseline of 32.3%. Those are their numbers. The stack is a Python process with a disk: it scales up, and idle it is still running — or the container sleeps and the graph goes with the disk. This agent keeps the same ideas in D1, R2, and Vectorize, and the Worker exits between questions.