Map who knows what across a GitHub org and trace past decisions

Percy Magombedze built Repo Brain solo for a FalkorDB graph hackathon and pointed it at twelve months of the public FalkorDB org: seven repos, about 3,100 pull requests, 2,200 issues, 25,000 comments and 360 people, loaded from the GitHub API into a graph without any model involved. Two agents built on Gemini function calling then work on it. A Curator agent picked the 250 most debated threads, read each one with its comments and wrote out the decision, the rationale, who made it and which comments state it. An Analyst agent answers questions such as who knows a folder, what breaks when a file changes, or why something was decided, but only through a fixed set of read-only graph tools, never by writing queries itself. Analyst workers also claim triage tasks the Curator queues for new issues. He used Claude Code as a coding assistant while building it.

The Curator produced 290 decisions from 206 of the 250 threads, tagged across 524 topics. Every answer the Analyst gives has to cite nodes it actually found; any name or file it cannot back up is marked unverified, and the evidence graph is drawn next to the answer. The author points out one miss: a pull request the Curator marked as only proposed had in fact shipped in a release, so the status is the model's reading while the release link is a fact. Path questions run in 8 to 140 ms over about 40,000 nodes. The demo runs on Railway in about 500 MB of RAM.

It is a hackathon project, run on one org by its builder, with no account of a team using it day to day. The public demo shares one free-tier Gemini key, so it is capped at 8 questions per visitor every 10 minutes. The author keeps decision provenance honest by dropping any comment id the model cites that was not in the thread it was shown.