GBrain: Garry Tan's OpenClaw Memory System (And Whether You Need It)

Garry Tan — president and CEO of Y Combinator — just open-sourced the memory system powering his personal OpenClaw and Hermes agent deployments. It's called GBrain, and it's one of the more impressive agent memory setups published publicly. Here's what it does and how to think about it for your own setup.

What GBrain Is

GBrain is a structured memory and knowledge brain for AI agents. Garry describes the production numbers that motivated it:

"17,888 pages, 4,383 people, 723 companies, 21 cron jobs running autonomously, built in 12 days. The agent ingests meetings, emails, tweets, voice calls, and original ideas while you sleep."

In practical terms: GBrain is a local database + skill system that lets your OpenClaw agent index, search, and enrich your personal knowledge base — contacts, companies, notes, meetings, ideas — using hybrid semantic search (vector + keyword).

Key technical specs:

• Database: PGLite (embedded Postgres, no server required) — ready in 2 seconds
• Search: Hybrid search (vector + BM25), results in ~0.12s
• Skills: 25 pre-built skills covering people enrichment, company research, meeting ingestion, idea capture, overnight memory consolidation, and citation repair
• Install time: ~30 minutes (agent-driven setup)
• Runtime: Bun

What It Actually Does

The core loop GBrain enables:

1. Your agent ingests raw inputs — meeting notes, emails, voice call transcripts, tweets, ideas — and stores them in the local PGLite database
2. Every person or company the agent encounters gets automatically enriched (LinkedIn, Clearbit-style data pulls)
3. Overnight cron jobs consolidate memory, fix broken citations, and surface cross-connections between entries
4. You query it conversationally: "what themes show up across my notes?" or "what do I know about [person]?"

The hybrid search means results are genuinely good — not just keyword matches. The example from the repo: searching for a PG essay concept returns results ranked by semantic similarity (0.94, 0.87, 0.81) across 12ms.

How Installation Works

GBrain is designed to be installed by your agent, not by you. The entire install flow is:

Retrieve and follow the instructions at:
https://raw.githubusercontent.com/garrytan/gbrain/master/INSTALL_FOR_AGENTS.md

Paste that into your OpenClaw agent. It clones the repo, installs GBrain, sets up the database, loads all 25 skills, and configures the recurring cron jobs. You answer a few questions about API keys. That's it.

Manual install is also documented for those who prefer it:

git clone https://github.com/garrytan/gbrain.git
cd gbrain
bun install && bun link
gbrain init       # local brain, ready in 2 seconds
gbrain import ~/notes/  # index your markdown
gbrain query "what themes show up across my notes?"

GBrain vs. OpenClaw's Built-In Memory

OpenClaw has its own memory system — memory.md, the workspace compiler, and optional LanceDB/vector memory. GBrain sits on top of that and adds structured relational data (people, companies, meetings as actual records), hybrid search, and automated enrichment pipelines.

Who Should Install GBrain

Good fit if you:

• Deal with a large network of people and companies (investors, sales, recruiting, partnerships)
• Want your agent to remember people across sessions and auto-enrich contact context
• Are ingesting lots of unstructured inputs — calls, meetings, notes — and want semantic search across them
• Already have OpenClaw running well and want to extend its memory capabilities

Skip it for now if:

• You're still getting your basic OpenClaw setup working — don't add complexity before the foundation is solid
• Your memory needs are simple (task tracking, decisions, project notes) — memory.md + heartbeat handles that fine
• You don't have Bun installed and don't want another runtime dependency

The Bigger Picture

GBrain is notable beyond the feature list. It's the first time a major tech figure has published the actual production agent brain they run daily — with real numbers and a working install path. The "21 cron jobs running autonomously" detail signals where serious agent operators are heading: not chat sessions, but persistent autonomous systems doing continuous background work.

That's the same direction ClawReady setups are built for. If you're using OpenClaw for serious business operations, GBrain is worth watching — even if you don't install it today.

GitHub: garrytan/gbrain