YC President Garry Tan Open-Sources GBrain, a Personal Memory System for AI Agents

Y Combinator president Garry Tan has open-sourced GBrain, a personal knowledge management system designed to give AI agents long-term memory. Released under the MIT license on April 10, 2026, the project amassed over 5,400 GitHub stars in its first 24 hours and reached more than 1.5 million people on X.

Key Highlights

• Persistent agent memory built on a markdown repository with Postgres/pgvector backend
• Hybrid search combining vector embeddings, keyword matching, and reciprocal rank fusion across 10,000+ files
• 37 operations including knowledge ingestion, entity detection, and multi-client access via CLI, MCP server, or HTTP
• MIT licensed with support for OpenClaw, Hermes Agent, Claude, Cursor, and Windsurf

How GBrain Works

At its core, GBrain stores knowledge as markdown files in a git repository. Each page follows a "compiled truth + timeline" pattern: a summary section followed by chronological entries documenting when information was learned or updated. The system indexes these pages into a Postgres database with pgvector for hybrid search.

Installation takes seconds using PGLite, an embedded Postgres 17.5 instance running via WebAssembly. No external server is required. For production scale, developers can connect to Supabase or self-hosted Postgres.

Tan described the vision on X: "If you want your OpenClaw or Hermes Agent to be able to have perfect total recall of all 10,000+ markdown files, GBrain is here to help." The system includes integration recipes for Gmail, Google Calendar, Twilio voice calls, and meeting transcripts.

The Dream Cycle Concept

One of GBrain's most talked-about features is "dream cycles" — autonomous overnight processing where agents enrich and consolidate knowledge while the user sleeps. Tan claimed the brain "is smarter than when I went to sleep" after these cycles run, challenging the stateless paradigm of most AI chatbots.

Independent Review Raises Questions

However, the launch was not without controversy. An independent code review published on DEV Community examined GBrain's three flagship features: compiled truth rewriting, dream cycles, and entity detection. The reviewers found that all three are implemented as markdown instruction documents that guide AI agents on what to do, rather than as executable code.

The review noted the absence of keywords like "rewrite," "schedule," "setInterval," and "timer" in the source files, concluding that the codebase contains "no rewrite logic, no scheduling mechanisms, no entity detection implementation." The MCP server integration was also flagged with twelve critical issues including race conditions and NULL embedding overwrites.

That said, the reviewers acknowledged that GBrain does include "reasonably competent infrastructure" — a PostgreSQL storage layer with pgvector, hybrid search using reciprocal rank fusion, and a chunking pipeline.

Why It Matters

The release reflects a growing movement toward giving AI agents persistent, user-controlled memory rather than relying on vendor-managed context windows. As AI agents become more prevalent in daily workflows, the ability to accumulate and retrieve knowledge across sessions could become a key differentiator.

Tan's position as YC president gives the project outsized visibility among startup founders who prefer customizable, self-hosted infrastructure over locked vendor ecosystems. Whether GBrain's agent-instruction approach to complex features proves effective or problematic will likely depend on how capable the underlying AI models become at following those instructions reliably.

What's Next

Tan has announced that full MCP support with Supabase Auth is coming soon, along with continued improvements to the integration ecosystem. The project remains actively maintained on GitHub.

Source: garrytan/gbrain on GitHub