We measure it. These are the latest scores from the Coppermind retrieval benchmark, measured 2026-06-25.
When you ask Coppermind about something you stored — a budget decision, a stakeholder preference, what you promised on the last call — recall@10 measures how often the right memory shows up in the top ten results. A score of 100% means that on this benchmark, the memory you needed was there 100% of the time.
The Coppermind retrieval benchmark is our internal, 9-query retrieval benchmark run on sanitized fixture data (no customer data is ever used). It covers exact-keyword lookups, semantic paraphrases (asking in different words than you stored), and knowledge updates — the hard case where a fact changed and the stale version must not win.
These are measured scores on a named internal benchmark at a point in time — not a universal guarantee. We re-run the benchmark as the retrieval pipeline evolves, and this page is regenerated from the committed baseline on every build, so the number you see is the number we measured.