labs · arc

cognee mirror-pilot

Can a memory substrate be trusted? A five-piece investigation of cognee — from the gotchas a builder hits to the question a founder actually needs answered.

Five pieces · June 2026

We set out to use cognee — a vector-plus-graph memory library — as the substrate for a recognition-style evaluation, with two corpora that had to stay isolated from each other. Proving that isolation, and then proving the evaluation method itself, turned out to be the real work.

This arc is the record of that pilot: the builder-level gotchas that almost let one corpus bleed into another, the founder-level question underneath ("can I trust what this thing remembers?"), the methodology we built to answer it, and the principles that fell out. Read in order; the questions deepen piece to piece.

Part 1 · June 2026

Cognee 1.0 dataset isolation — this post is held for correction

Held for correction. This post claimed cognee 1.0's dataset isolation is off by default; that was verified wrong at source on 2026-07-17 — the leak we described came from our own config, not a cognee default. Kept public with an erratum while the rewrite is worked.

Forthcoming in this arc: the founder's question, the recognition-not-diagnosis methodology, the integrated flagship read, and seven principles for trusting a memory substrate. Each lands here as its mirror publishes.