GBRAIN Deep Dive: The AI Knowledge Engine Built by YC’s CEO

What is This Project

GBrain is an open-source personal knowledge management system created by Garry Tan, the current CEO of Y Combinator. Project URL: github.com/garrytan/gbrain. Its positioning can be summed up in one sentence: Search gives you raw pages; GBrain gives you answers.

Traditional note-taking tools and knowledge bases are essentially about “storing” and “searching” — you store a thousand notes, keyword-search when needed, then read through them one by one and manually piece things together. GBrain adds a synthesis layer on top: it doesn’t just help you find relevant documents; it understands relationships between entities across documents, directly outputs a complete answer with cited sources, and proactively tells you “here’s what’s missing from your knowledge base on this topic.”

The project has garnered over 18,600 Stars on GitHub, is open-sourced under the MIT license, written in TypeScript, with core code accounting for 96.8% of the repository.

Core Capabilities Breakdown

Knowledge Synthesis Engine

GBrain’s key differentiator lies in its two query modes:

• gbrain search: Traditional retrieval mode, returns raw pages sorted by hybrid scoring. Fast, no LLM tokens consumed.
• gbrain think: Synthesis mode. Generates a structured answer based on search results, complete with inline citations and gap analysis.

The think mode is GBrain’s real selling point. It doesn’t simply dump retrieved documents into an LLM for summarization. Instead, it traverses the knowledge graph to find relational paths between entities, then generates answers based on those connections. This means even if two notes were never reviewed together, GBrain can discover the link between them.

Automatic Knowledge Graph Construction

Every time you write a piece of content into GBrain, the system automatically extracts entity references (names of people, companies, concepts, etc.) and creates typed relationship edges. The key point: this process does not require LLM calls — it’s purely done through a rules engine, at zero cost.

This solves a long-standing pain point in the knowledge graph space: manual maintenance is too tedious, and LLM extraction is too expensive. GBrain takes a pragmatic middle path — extract entities with rules, store relationships in a graph, and reserve LLM power for the final synthesis stage.

Hybrid Search Strategy

GBrain’s retrieval layer blends multiple search techniques:

• Vector search: Matches based on semantic similarity, understanding “related meaning.”
• BM25 keyword matching: Traditional exact keyword retrieval ensures no explicitly mentioned content is missed.
• Reciprocal Rank Fusion: Merges rankings from multiple retrieval methods.
• Source-Tier Boost: Allows different weights for content from different sources.
• Intent-aware query rewriting: Understands what you really want to ask, not just literal text.

According to published evaluation data, on a 240-page assessment corpus, GBrain achieves a P@5 (precision at top 5 results) of 49.1% and an R@5 (recall at top 5) of 97.9%. Compared to a version without the knowledge graph, P@5 is improved by 31.4 percentage points.

Technical Architecture

Data Storage: Markdown + Git

GBrain’s data layer design is distinctive: all knowledge is stored as Markdown files in a Git repository. This means:

• Your knowledge base inherently has version control.
• You can edit files directly with any text editor.
• Even if GBrain stops being maintained, your data remains readable plain text.
• Multi-device sync is simply git push / git pull.

Dual Database Engine

• PGLite (default): Embedded Postgres 17 running via WASM, zero configuration, suitable for personal use, supports about 50,000 pages.
• Postgres + pgvector: Suitable for teams and large-scale deployments, can use Supabase or self-hosted instances.

Organization Model

GBrain organizes knowledge along two orthogonal dimensions:

• Brain: A database instance, corresponding to an independent knowledge space.
• Source: A sub-repository within a Brain, e.g., wiki, note collection, knowledge base.

Processing Loop

The entire system workflow is a closed loop:

Signal Capture → Search & Retrieval → Answer Generation → Write Back to Knowledge Base → Auto-establish Links → Sync

This means that while GBrain answers your questions, it is also continuously enriching its own knowledge graph.

Task Queue (Minions)

GBrain includes a built-in persistent task queue based on Postgres, modeled after BullMQ’s design. Its purpose is to run “sub-agents” — LLM tool-calling loops that can survive crash recovery. A two-phase (pending → done) persistence mechanism ensures tasks are not lost.

SCHEMA System

GBrain does not force a fixed knowledge classification scheme. Instead, it provides a customizable Schema system:

• gbrain-base: Default classification scheme includes eight directories: people/, companies/, concepts/, meetings/, deal/, daily/, originals/, writing/.
• gbrain-recommended: Extends the base with 13 additional directories.

You can also define your own Schema Pack. The project provides a complete set of Schema management commands:

gbrain schema active       # View currently active Schema
gbrain schema list         # List all available Schema Packs
gbrain schema detect       # Auto-detect appropriate classification for content
gbrain schema suggest      # Recommend Schema optimization suggestions
gbrain schema use my-pack  # Switch to a specified Schema

Data Ingestion

GBrain covers almost all information capture scenarios:

gbrain capture "A piece of text you want to remember"           # Direct input
gbrain capture --file ./notes/today.md                          # Import from file
echo "from a pipe" | gbrain capture --stdin                     # Pipe input

Beyond the command line, it also supports:

• Webhook ingestion: POST to /ingest endpoint with Bearer Token authentication.
• iOS Shortcuts / Drafts: Quick capture from mobile.
• Voice input: Via Twilio + OpenAI Realtime, or a custom STT+LLM+TTS pipeline.
• Email and Calendar Webhooks: Automatically ingest schedule and email content.

Integration Ecosystem

Embedding Providers

GBrain offers configuration for 16 embedding models, covering:

MCP Client

GBrain can run as an MCP Server, integrating with mainstream AI tools:

gbrain serve          # stdio mode
gbrain serve --http   # HTTP mode, supports OAuth 2.1

Supported clients include Claude Code, Claude Desktop, Cursor, Windsurf, Perplexity, ChatGPT.

Prebuilt Skills

The project includes 43 built-in Skills, covering signal capture, content ingestion, knowledge enhancement, query retrieval, operations, citation fixing, daily journal management, scheduled tasks, evaluation frameworks, and more.

GARRY TAN’s Production Instance

According to the README, Garry Tan’s own GBrain instance has reached a substantial scale:

• 146,646 pages indexed
• 24,585 people tracked
• 5,339 companies tracked
• 66 scheduled tasks running automatically

As the CEO of YC, Garry Tan interacts with a huge number of founders, companies, and deal information daily. GBrain is essentially his self-built “investor brain external drive” — turning fragmented contacts, companies, deals, and meeting notes into a structured system that allows cross-dimensional querying and analysis at any time.

Who Should Use It

Heavy knowledge workers: Investors, researchers, analysts, journalists — anyone who needs to quickly find connections and generate insights from massive amounts of fragmented information. GBrain’s knowledge graph and synthesis capabilities are designed precisely for these scenarios.

Developers and technical teams: Those who want to build a programmable, extensible private knowledge retrieval system deeply integrated with existing AI toolchains. GBrain’s MCP Server mode, 43 prebuilt Skills, and extensive embedding model support make it a flexible knowledge infrastructure.

Users who value data sovereignty: All data is stored as Markdown in a local Git repository, with no dependency on any third-party cloud service. Even if you stop using GBrain, your data remains fully readable.

Summary

GBrain represents a new direction in personal knowledge management: not just storage and retrieval, but understanding and synthesis. It integrates knowledge graphs, hybrid search, and LLM synthesis into a single stack, while adhering to the engineering philosophy of “data is files, files are Git.” It strikes a balance between openness and practicality.

Given Garry Tan’s own production use with over 140,000 pages, this is not a proof-of-concept project — it’s a system battle-tested under real, high-intensity usage.

Published publicly. Please cite the source when republishing.