inherit-brain

inherit-brain is a company brain for the product you just inherited. It turns a team's Slack messages, meeting transcripts, documents, pull requests and bug histories into one FalkorDB graph per product, and an agent answers questions by calling nine read-only Cypher tools. Every answer shows its tool calls, the Cypher they ran and the records they returned, so you can check the reasoning instead of trusting it.

The headline result is measured, not claimed: on 1,203 held-out questions from Salesforce's HERB benchmark (a simulated company with 30 products), the graph tools scored 80.0 against 39.2 for HERB's own retrieval tools in the same loop and 19.2 for hybrid retrieval, using the same model. Swapping the graph tools for retrieval costs 40.8 points. The explorer site shows all 1,203 questions with gold answers, per-method scores and tool traces, including the questions inherit-brain got wrong, and it makes no model calls of its own.

Vibe-coded evidence, in the author's own words on DEV (18 October 2026): "I built this with Claude Code and Codex as coding tools; the benchmark choice, experiment design, gates and release decisions were mine." The plan and hashes were posted publicly before any held-out question was answered, and the full results, including a what-if page for what happens if an employee leaves, are on the demo site. GitHub stars: 0 at the 28 October 2026 check (github.com/zeyusuntt/inherit-brain).

Honest limits: the author reports the lead is smaller on never-seen question wordings (+20.1) than on wordings shared with development questions (+57.9), that 46 of 659 answers to answerable questions cited no record, and that the judge was the same open model, so content scores are not comparable with the HERB paper's GPT-4o-judged numbers. The demo is a recorded explorer over a simulated company, not your own codebase; running it on your records needs the local FalkorDB + Docker setup in the repo.