GBrain for OpenClaw: Garry Tan's AI Agent Memory System Explained
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
This video is about GBrain, an AI agent memory system from Garry Tan, the president of Y Combinator. If your OpenClaw or Hermes Agent keeps forgetting who you are and what you're working on, GBrain is built to fix that. I explain what it does, how the setup works, and what you need before you install it. I recorded this in April 2026. Tools like this change fast, so check the current docs in the repo before you set it up.
What I cover in this video
- 0:00: What GBrain is and why it got more than 3,700 GitHub stars in under 10 hours
- 0:45: The core memory problem with AI agents
- 1:30: How GBrain builds a searchable knowledge base from your life
- 3:01: How Garry Tan uses it with 10,000+ markdown files and 13 years of calendar data
- 3:47: The technical side: local PGLite database, one-block install, hybrid search, and recipe files
- 5:20: Why this changes how you use an agent day to day
- 8:21: What you need to get started, including which models to use
- 10:36: My current memory setup and why I want to try GBrain
The problem: your agent doesn't know you
Agents are really capable. They can write code, research topics, answer questions, and build full web and mobile apps. But they know nothing about you: your business, your context, your history, your relationships. Every conversation starts from zero, so you keep re-explaining yourself. "I'm a founder, I sell this product, my audience is this, here's what I've been working on." Doing that every day gets exhausting.
Most people handle this with a memory file, but that's manual. You have to decide what to save and write it down yourself. If you forget, it's gone.
What GBrain actually does
GBrain is an opinionated brain setup for AI agents. It creates a searchable knowledge base from everything in your life: files, meetings, emails, tweets, calendar events, voice calls, and the random ideas you capture. Two things make it work:
- Your agent reads from it before every single response. Not just some of the time. Every time.
- After every conversation, it writes back to it. So the system keeps picking up more context about you, your work, and your life.
The agent doesn't just know things. It learns, and it gets a little smarter about you every day.
Garry Tan runs more than 10,000 markdown files through GBrain. That includes about 3,000 people dossiers with notes and contact details, plus 13 years of calendar data. When his agent answers something, it's pulling from that whole history.
Here's what that looks like in practice. Say you have a call with an investor you met two years ago. With a normal agent, you'd dig up your notes, paste them in, and ask for help prepping. With GBrain, the agent already has the notes from your last three conversations with that person, along with what you discussed and what you promised. You just say "help me prep for this call with this person" and get useful context back with no copy-pasting.
How the setup works
The technical side is simpler than you'd expect:
- About 30 minutes to set up.
- No server required. GBrain runs on PGLite, an embedded Postgres database that lives on your machine. Your data stays local and doesn't go to a cloud service.
- Works with OpenClaw and Hermes Agent. On OpenClaw, you paste in one block of config and the agent installs itself. You don't have to set up tables or write queries.
- Hybrid search. It combines vector embeddings with regular keyword search. Vector search finds things by meaning, not exact words. If you wrote "need to follow up with Brett about the call scheduling" six months ago and you ask your agent about Brett now, it finds that note even if the wording doesn't match.
- Recipe files for integrations. Voice, email, calendar, Twitter, and more. You pick the recipes you want, drop them in, and that data starts flowing into the knowledge base.
What you need to get started
- OpenClaw or Hermes Agent already running. GBrain plugs into one of them. If you already use OpenClaw, you're halfway there.
- A capable model: Claude Opus 4.6 or GPT-5.4. Hybrid search and context synthesis need a strong model to be useful. You can try a lighter model, but quality drops fast.
- The GBrain repo. Pull the config, pick your recipes based on the data you want coming in, and paste the install block into your agent. The agent handles the rest of the setup itself.
- Feed it data. Add your notes, your calendar, and any existing files you want it to know about. The more you add at the start, the sooner it becomes useful.
You don't need 13 years of data like Garry has. Even a few hundred files and a few months of calendar data will change how your agent responds.
Why this matters for how you use agents
Most people use their agent like a really smart search engine. They ask, it answers, and that's using maybe 20% of what it can do. GBrain turns the agent into the memory keeper. You do your work and have your calls, and the system builds context in the background. Then when you open your agent, it already knows things.
That's the step from one-off tasks to an agent that can be proactive. For example: "Based on your history with this client and the notes from your last three calls, here's what I think you should do." That only works with a lot of context. In my view, the thing that separates a useful agent from a transformational one isn't the model or the tools. It's whether the agent knows who it's working for.
The fast GitHub reaction makes sense to me because this is Garry's real setup, not a demo. He's running it on a large amount of his own data.
My own memory setup right now uses supermemory.ai, where my agents push work and memories into their own cloud container. I also use Neon as a backup database for unstructured data, plus Lossless Claw. In the video I say I plan to set up GBrain on my OpenClaw and get it running on Hermes Agent too. I want to see how it feels when my agent pulls from a structured knowledge base instead of flat memory files. If you're building out AI agents you plan to run for a long time, it's worth 30 minutes to try.
Key takeaways
- GBrain is Garry Tan's open-source memory system. It reads from your knowledge base before every response and writes back after every conversation.
- It runs locally on PGLite (embedded Postgres), so you don't need a server and your data stays on your machine.
- Hybrid search (vector plus keyword) finds notes by meaning, not just exact wording.
- Recipe files bring in voice, email, calendar, Twitter, and other data.
- You need OpenClaw or Hermes Agent plus a strong model like Claude Opus 4.6 or GPT-5.4. Setup takes about 30 minutes.
FAQ
What is GBrain?
GBrain is an open-source AI agent memory system from Garry Tan, the president of Y Combinator. It builds a searchable knowledge base from your files, meetings, emails, tweets, calendar, and voice calls. Your agent reads from it before every response and writes back to it after every conversation.
Does GBrain work with OpenClaw and Hermes Agent?
Yes. GBrain works with both OpenClaw and Hermes Agent. On OpenClaw, you paste one block of config and the agent installs itself.
Does GBrain store my data in the cloud?
No. GBrain runs on PGLite, an embedded Postgres database that lives on your machine. You don't need a server and your data stays local.
What model do I need to run GBrain?
I recommend Claude Opus 4.6 or GPT-5.4. Hybrid search and context synthesis need a capable model, and quality drops fast on cheaper ones.
Transcript
Read the full transcript
0:00If you're having memory issues with your open claw, then this is going to be the video for you. I'm going to be talking about a repo called G brain by Gary Tan. So, it just dropped about a few days ago and the president of Y Combinator has already have over 3,700 stars and counting. Like I said, I have it linked down below and I want to explain why this thing is spreading so fast because once you understand what it actually does, and you'll get it immediately. So, here's the core problem most people have with just AI gen AI agents in general is memory. Your agent is smart, like actually really smart. You have super intelligence at the tip of your fingers, but the main thing that happens is they forget and you feel like it has dementia. So, this fixes that. You know,
0:45it can write code, it can research topics, it can even answer questions, it can build full-on web apps and mobile apps, but it knows nothing about you. Like nothing about your business, your context, your history, your relationships. Every single conversation you start is basically starting from zero again. You have to re-explain yourself constantly. You know, like I'm a founder, I sell this product, I have this many subscribers on my YouTube channel, my audience is this, here's what I've been working on. It seems like you have to do that over and over again on a daily basis and it's exhausting. G brain fixes that permanently. Here's what it actually is. G brain is an opinionated AI agent brain setup. What it does is create a searchable knowledge
1:30base from everything in your life. Your files, your meetings, your emails, your tweets, your calendar events, voice calls, random ideas you've captured cuz I know you got those. I got them. Everything goes in and then your agent reads from it before every single response. Not just sometimes, every time. And after every conversation, it writes back to it. So, the system is constantly getting more context about you, your work, your life. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watching videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full open claw course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually
2:15set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or open claw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. The agent doesn't know things, it learns. Every day it gets a little smarter about who you are, which is pretty cool. But what makes this thing more crazy is who built it and how he's
3:01actually using it. Gary Tan runs 10,000 plus markdown files through through G brain. He has 3,000 people dossiers in there with notes, contacts on 3,000 different people and 13 years of calendar data. 13 years. That's a decade plus of context about every meeting, every project, every relationship that matters in his life. When his agent answers something, it's pulling from that entire history. That's not a smart AI assistant anymore, that's literally a second brain that knows you. Think about what that actually means in practice. Say you're getting on a call with an investor who you met 2 years ago. With a normal AI agent, you'd go find the notes yourself, paste them in, and ask your agent to help you prep. With G brain, your agent already knows this person. It
3:47already has the notes from your last three interactions. It already knows what you discussed, what you promised, what the relationship looks like. You just say, "Help me prep for this call with this person." and it comes back with actually useful contexts. No copy-paste, no manual lookup, it just knows already. That's the shift from tool to collaborator. Let me walk you through the technical side real quick because it's simpler than you think. The setup takes about 30 minutes. No server required, G brain runs on a PG light database, which is embedded in Postgres that lives right on your machine. So, this is not a cloud thing where your data is going somewhere. It's local, it's yours. It works with open claw and Hermes agent. If you're on open claw like I am, you literally paste in one block of config and the agent installs
4:34itself. It does all the work. You're not manually setting up tables or writing queries or anything like that. The search is hybrid search. That means it's combining your vector embeddings with regular keyword search. Why does that even matter? Because vector search finds things by meaning, not just exact words. So, if you wrote, "Need to follow up with Brett about the call scheduling." 6 months ago and now you ask your agent about Brett, it finds it even if the words don't match exactly. And then all the integrations come via recipe files. Voice, email, calendar, Twitter, all of it. You pick the recipes and you want you pick the recipe you want, drop them in, and the data starts flowing into the knowledge base. So, now I want to talk about what this actually means for how
5:20you're going to use an AI agent because I think a lot of people are underestimating this. Right now, most people interact with their agent like it's a really smart search engine. It's cool, but you're using like 20% of the capability. You ask it for something, it answers. It's that con type of conversation over and over. Maybe you have some stuff in a memory file, but it's manual. You have to decide what to save, you have to write it down, and if you forget to, it's gone forever. G brain flips this. The agent becomes the memory keeper. You don't have to think about it, you just have to live your life, do your work, have your calls, and the system is building context about you automatically. Then when you go to actually use your agent, it shows up already knowing things, already carrying that context. You know how the best assistant you've ever had, the one who
6:05actually got things done. They remembered everything. They remembered that you mentioned a problem 3 weeks ago and they already worked on it. They remembered who do you talk to and what you said. That's what this is trying to do. Not by being smarter than other AI, but by actually having the context to work with. Now, let me tell you why 3,700 stars in 10 hours is significant. Most GitHub repos that get that kind of reaction are from big companies or well-known developer tools. Gary Tan posting a personal infrastructure project is different. This is the president of one of the most influential startup accelerators in the world saying, "Here is how I actually use AI agents. This is my real setup. Not a demo, not a blog post about hey how AI is transforming things." His actual system. That matters because it's
6:50validated. This isn't just like a theoretical side project. It's not some guy saying, "Here's what I think would be a cool demo." It's a guy with massive amounts of real data and real context running this in production and saying you should use it, too. So, how do you actually get started with this? You need either open claw or Hermes agent setup already going. G brain plugs into one of those. If you're struggling to keep up with content, what I'm about to save you about 40 days worth of work. I built something called Content Machine. It's 10 AI agents that run on the open claw orchestration and they handle everything. Scripts, thumbnails, X posts, [music] blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers
7:36on YouTube in 7 days using this exact system. Every single morning I wake [music] up and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them and I move on with my day. It works for any niche, fitness, [music] finance, real estate, marketing, whatever you are building and it is 100% completely customizable to your use case. So, you get [music] the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns how you talk, and it writes so it doesn't sound like AI slop. $97 one time, it's not a subscription. I'll put the link down below and you'll thank me
8:21later. If you're already running on open claw, you're already halfway there. You haven't set up open claw yet, I've got videos on that, just go check it out. The second thing is you need Claude Opus 4 6 or GPT 5 4. This is not a setup you can run on cheaper models. The hybrid search and context synthesis requires a capable model to actually be useful. You could try to run it on something lighter, but the quality will drop super fast. The third thing is you head to the G brain repo, which I'll link to in the description, you pull the config, you pick up your recipe based on what data you want it flowing in, and you paste the install block into the agent setup. The agent walks itself through the rest of the setup. It's really easy. Then you start feeding it your notes, your calendar, any existing files you want it to know about. The more you put in up
9:06front, the faster it becomes more useful. Gary's been building his for years, so obviously his is pretty crazy. But even starting with a few hundred files, a few months of calendar data, you'll feel the difference in how your agent responds. Here's the thing that I keep coming back to. Most people are using AI agents for one-off tasks. Write this email, summarize this document, answer this question. And that works, I agree. But the real unlock is when your agent knows you well enough to be proactive. When it can say, "Hey, based on your history with this client and the notes from your last three calls, here's what I think you should do." That level of usefulness requires context, a lot of it. And G brain is the most practical system I've seen for building that context up over time. There's a reason the people who've been using personal AI agents the longest all talk about memory
9:51the same way. It's the thing, right? Not the model you pick, not the tools you to The thing that separates a useful agent from a transformational one is whether you it actually knows who it's working for and knows more about you. G brain is Garry Tan's answer to that problem. And given that he runs one of the most active startup accelerators on the planet and is presumed and is presumably talking to hundreds of founders every year, his agent needs to carry a lot of context. The fact that he built this, uses itself uses it himself and then open source the whole thing is pretty cool. I'm going to be setting this up on my own open claw here today and I'm getting my own Hermes agent set up here in the next day. So, I'm going to be playing around with this. I already have memory set up in a way that I've dialed
10:36in. We have things going to supermemory.ai where the agents push a lot of their work and memories in their own container in the cloud. I also have neon as a backup database where it pushes unstructured data. We also have lossless claw. I want to see what it actually feels like when my agent is pulling from a real structured knowledge base instead of just flat memory files. I'll report back and make a video later to follow up. But if you're serious about building an AI agent setup that actually works long-term, not just something that does tasks on demand, but something that gets more useful the longer you run it, G brain is worth your 30 minutes. Get it installed. It's Sunday. You could get it done. I mean, what else are you doing today? I got a link down below. And if you want to learn how to build out AI agent setups and workflows and figure
11:22out get getting these types of tools installed, it's exactly what we walk through in the community. We have 220 members and I launched this community 27 days ago. I'll put the link to it down below. We do nine boot camp calls a week. The goal is for you to come into one of these calls, get your question answered so you could take the next step on your AI journey. If you haven't subscribed to the channel, please do already cuz we release four to six videos every single day covering all AI news. We'll see you in the next video. Be blessed.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Context Overflow Explained: Shared Memory for AI Agents
Context Overflow is a Q&A knowledge-sharing layer for AI agents. Agents search past solutions before starting a task, ask questions when they get stuck, and share what they figure out, so the next session doesn't start from scratch. I also cover OpenClaw's memory wiki and active memory, plus GBrain.
OpenClaw Memory Wiki: Karpathy's LLM Knowledge Base Explained
OpenClaw 4.7's Memory Wiki uses the same ingest, compile, lint loop Andrej Karpathy described for LLM Knowledge Bases. It turns raw notes into structured, timestamped wiki pages, compiles digests your agents read, and flags stale or contradictory facts on a health dashboard.
Hermes Agent v0.15 Update: Kanban Swarms, Session Search, Security
Hermes Agent v0.15, the Velocity Release, adds a multi-agent Kanban board, rebuilt session search that's about 4,500 times faster and free, promptware defense, Bitwarden secrets support, and a big code refactor. OpenClaw builders should copy two things now: the job-board pattern and per-task model routing.
