I measured how much my coding agent actually knows about my stack. It was 40%.

<p>My coding agent answers every question about my stack with total confidence. Last month I finally asked the question that should have been first: <em>how much of that is true?</em></p> <h2> The experiment </h2> <p>I took the platform docs my agents rely on daily — a games platform's RGS API, its math SDK, the approval checklist — wrote down what an agent <em>should</em> be able to answer, generated ~30 control questions from that, and graded the agent's answers against the actual docs.</p> <p>It scored around 40%. The worst part wasn't the score — it was that every wrong answer <em>sounded exactly like every right one</em>. Same confidence, same fluency, same code blocks. The model's training data was simply older than the docs, and nothing in the conversation could tell me which answers were from 2024.</p> <h2> Context files don't fix this </h2> <p>The standard fix is a CLAUDE.md / skills folder / a memory MCP server. I had all three. Three problems survived:</p> <ol> <li> <strong>You pay for every word, every session</strong> — the whole file rides along whether the task needs it or not.</li> <li> <strong>They rot silently.</strong> A file written in March is confidently wrong by June, and nothing tells you.</li> <li> <strong>You can't measure them.</strong> Nobody knows what a folder of markdown actually covers until the agent fails in production.</li> </ol> <h2> So I made the knowledge sit an exam </h2> <p>I built <a href="https://mozg.sh" rel="noopener noreferrer">mozg</a> around one mechanism: every knowledge base ('brain') is scored against its own purpose.</p> <ul> <li>Paste one docs URL. The crawler finds every page (GitHub tree, llms.txt, sitemap, or a link walk) and extracts it into atomic, searchable notes.</li> <li>The goal you wrote becomes ~30 exam questions. The brain sits them after every ingest. 'Trained 87%' is a measured number, and the failed questions are listed right on the brain's page.</li> <li>The exam deliberately asks about things the brain does NOT cover yet. The failures are the point — they tell you exactly what to feed it next.</li> </ul> <p>Agents connect over MCP (one command in Claude Code / Codex / Cursor) and search server-side — the context cost of an answer is the three notes it actually needed, not the 700 the brain holds.</p> <h2> The part I didn't plan: it learns from being used </h2> <p>Once real agents were querying brains, the logs contained something better than any test I could write: <strong>the questions the brain failed to answer.</strong> Now every search that returns nothing becomes an exam question automatically, and the next re-read of the sources goes hunting for it. Corrections agents file become owner-reviewed notes. The tenth user gets a measurably better brain than the first.</p> <p>That same exam turned out to be a curriculum: <a href="https://learn.mozg.sh" rel="noopener noreferrer">learn.mozg.sh</a> serves any brain as a spaced-repetition course — read, recall, quiz — with the brain's exam as the final. The scoreboard shows your percentage next to your agent's. Beating your own agent is weirdly motivating.</p> <h2> Honest limitations </h2> <ul> <li>The judge is a model, so scores wobble ±3-4 points (majority voting tames most of it).</li> <li>A brain is only as good as its sources — the exam tells you <em>that</em> material is missing, not where to find it.</li> <li>For stable, well-known knowledge (Python stdlib), a brain adds nothing — models know it. Brains earn their keep where docs move faster than training cutoffs.</li> </ul> <h2> Try it </h2> <p>The catalogue has free, no-card brains for exactly those fast-moving stacks: Next.js App Router, Expo/React Native, Svelte 5, Tailwind v4, the MCP spec itself. Connect one to your agent in a minute: <a href="https://mozg.sh" rel="noopener noreferrer">https://mozg.sh</a> — or study one yourself at <a href="https://learn.mozg.sh" rel="noopener noreferre