正文 Markdown
AI Heroes ran gbrain against their own production memory stack on 150 real questions from their actual corpus. gbrain won 58 head to head matchups. Their own system won 7. The methodology, the surprises, and what the aggregate score hides 🧵 TL;DR: → 352 file corpus, 150 questions built from real operator sessions, not synthetic evals → gbrain won 58 questions vs qmd's 7 in apples to apples retrieval, an 8.3x win ratio → gbrain ran 41x faster: 608ms median vs 25,138ms for qmd native pipeline → qmd's LLM reranker actively reduced recall on this corpus → gbrain's graph extractor produced 0 typed links, every win came from hybrid retrieval alone Marco and the AI Heroes team disclosed the conflict of interest directly: they run qmd in production and had a vested interest in it winning. The data did not cooperate. They are pulling the reranker from production qmd regardless of whether they migrate to gbrain. That is the kind of benchmark worth reading. @garrytan @openclaw