GBrain: Complete Build Specification

Garry Tan’s (YC CEO) complete build spec for GBrain — an open-source personal knowledge management system using a single SQLite database.

Architecture

• Storage: Single brain.db SQLite file with FTS5 + vector embeddings
• Knowledge Model: “Compiled truth + timeline” — current intelligence above the line, append-only evidence below
• Intelligence Layer: Fat markdown skill files guide AI agents
• CLI: gbrain get/put/search/query/import/export/serve
• MCP Server: 14 tools for Claude Code / Cursor integration

8 Core Tables

pages, page_fts, page_embeddings, links, tags, raw_data, timeline_entries, ingest_log

”Compiled Truth” Pattern

Each page splits into:

1. Above the line: Current best understanding, actively updated
2. Below the line: Evidence timeline, append-only

Similar to intelligence analysis — conclusions separated from evidence.

Tech Stack

• Bun + native SQLite
• FTS5 (Porter stemmer, unicode61 tokenizer)
• OpenAI text-embedding-3-small for vectors
• Official MCP SDK

Migration

Designed to import Garry’s existing 7,471-page wiki. Lossless, round-trippable. ~22,500 embedding API calls needed.

Key Distinction from Karpathy’s LLM Wiki

More engineering-focused: concrete schema, CLI, MCP integration, migration plan. Takes Karpathy’s conceptual pattern and turns it into a buildable product spec.