Battle of the Personal Brains · San Francisco

Cognee is already good. I wanted it to scale to zero.

Cognee’s own head-to-head, 24 HotPotQA questions, put their graph at 0.93 human-like correctness and 0.84 F1. On BEAM, at 100k tokens, they report 79% where a Llama-4 RAG baseline sat at 32.3%. Those are Cognee’s published numbers. I did not re-run them.

That memory lives in a Python process, with Kuzu or Neo4j and LanceDB on a disk. It scales up. Leave it idle and the process is still there, or the container sleeps and the disk goes with it. I wanted the same ideas to scale down to zero: no process between questions. The page stays Markdown in R2. The graph is D1 rows. Search is Vectorize plus the triplet score. A Workflow extracts, writes, and exits.

The agent is the Strands TypeScript SDK in the same isolate. A container starts only when a tool needs a shell. Ask it something and the graph is what it remembers.

Try the agent Memory graph GitHub · cloudnee GitHub · strands

IdleNothing provisioned between turns. The Worker returns. The Workflow is finished.

A fresh factD1 is the record on the next read, even while Vectorize is still catching up.

What I keptStable entity id, one extraction per chunk, triplet score, supersede in SQL, path permissions.

What this is notA claim that cloudnee beat Cognee’s benchmark. The nine jobs are a behavior demo.

Cognee sources: knowledge-graph memory benchmarks · behind the benchmark numbers. Cognee itself: topoteretes/cognee.