Garry Tan, President and CEO of Y Combinator, has open-sourced GBrain—a production brain layer for AI agents that he says runs his own OpenClaw and Hermes deployments across 146,646 pages, 24,585 people, and 5,339 companies. The project landed on Hacker News this week with the kind of confidence you expect from someone who runs the most influential startup accelerator in Silicon Valley: "It's the brain layer your AI agent has been missing—the only one that does synthesis, graph traversal, and gap analysis in one box."

What Makes GBrain Different

The core pitch is a familiar pain point with a fresh angle. Most personal-knowledge tools return chunks of raw text when you query them—"here are 10 pages mentioning your question," which leaves the actual synthesis work to you. GBrain's "think" command runs retrieval, then composes an answer across results with explicit citations and an honest gap analysis: what did it not find? What might be stale? The example in the README is telling: ask about a meeting with Alice tomorrow, and instead of dumping five pages, you get her role, your last conversation date, three open items between you, and a heads-up that nothing's been added to her profile in six weeks. That's synthesis doing the work, not just retrieval.

The Knowledge Graph Angle

The technical differentiator is self-wiring graph extraction that happens with zero LLM calls. Every page write extracts entity references and creates typed edges—works_at, invested_in, founded, attended, advises—with pattern matching on wiki-style [[links]] and structured syntax. Multi-hop traversal lets you ask questions like "who works at Acme AI?" or "what did Bob invest in this quarter?" that vector search alone can't reliably answer. The benchmarks back the claim: P@5 of 49.1% and R@5 of 97.9% on a 240-page Opus-generated corpus, with a +31.4 point P@5 lift over graph-disabled variants and vector-only RAG baselines.

Company Brain and Multi-User Access

Tan is positioning GBrain as both personal brain and company brain. The multi-user deployment scopes data by login—query results filter to what you're allowed to see, with Tan claiming zero leaks across every read path tested (search, list, lookup, multi-source reads). Each team member gets their own slice of institutional memory, which maps directly onto YC's recent "Request for Startups" focus on company brains. "If you're building in that space," he writes, "you might as well build on this."

Installation and Integration

The install target is explicitly AI agents first, humans second. The recommended path has your agent read INSTALL_FOR_AGENTS.md and execute it autonomously—Tan claims ~30 minutes end-to-end with no manual configuration beyond answering API key questions. PGLite handles the database layer in two seconds without a running server or Docker dependency. For existing coding agents like Claude Code, Cursor, Codex, or AlphaClaw, you paste the same instruction and let the agent handle it. GBrain exposes 30+ tools over MCP (stdio and HTTP), with OAuth 2.1 + PKCE required for remote clients like ChatGPT.

Schema Packs and Extensibility

Rather than enforcing a fixed taxonomy of notes/tags/people, GBrain ships bundled schema packs and lets you author custom ones. The default (gbrain-base-v2 as of v0.41.22) defines 15 canonical types—person, company, media, tweet, deal, email, project, note, and others—with subtypes pushed to frontmatter. Three CLI verbs (detect, suggest, review-candidates) let the brain infer your filesystem's actual shape and propose type mappings you approve before activation. Agents can evolve the schema themselves via 14 dedicated CLI verbs plus an MCP operation (schema_apply_mutations), with atomic file locks and audit logging.

Eval Framework Included

The project ships its eval framework alongside the core product. NamedThingBench hard-gates retrieval quality across title-substring, alias-synonym, generic-to-named, and multi-chunk-dilution families—regressions fail CI loudly. Suspected-contradictions surfaces conflicts between takes and facts using a query-conditioned LLM judge with persistent caching. Real queries can be captured (gbrain eval export) and replayed against code changes for regression testing. The LongMemEval public benchmark runs against the hybrid retrieval system.

Key Takeaways

- GBrain solves "search returns pages, not answers"—synthesis layer composes responses with citations and gap analysis - Self-wiring knowledge graph extracts typed edges (works_at, invested_in, etc.) without LLM calls on every write +31.4 P@5 lift over vector-only RAG suggests the graph layer is doing real work, not just decoration Install target is explicitly agents first: paste a URL, answer API key questions, ~30 minutes to working brain Multi-user company-brain mode ships with OAuth-scoped access and claims zero cross-user data leaks in fuzz testing Schema packs let you define your own taxonomy; agents can evolve it autonomously via MCP operations