Garry Tan's GBrain Dynamically Loads Relevant Context for AI Agents

I want GBrain to be the retrieval layer like Postgres for your agent. It's custom built to be the just-in-time 3 book librarian to give your AI agent wings.

It's MIT-licensed free open source with thousands of users using it today.

https://github.com/garrytan/gbrain

The fastest way to *try* GBrain is hosted at http://gbrain.io

Agreed, gbrain is a legit unlock for surfacing the right knowledge (have a Hermes agent running on it)

But a company isn’t a library of files.

You need a living map of the territory (entities, relationships, states, events) that agents can navigate.

The map is the new executable layer IMO.

PS. Building http://ortelian.com based on it

@garrytan big unlock, sure. but when the agent can load anything, it loads the wrong thing with total confidence. Retrieval precision is the bug that quietly costs you

@garrytan Thanks for Gbrain.

I built a "company brain" for my needs last week based on your concept.

Today it got its first customer. A client who uses WhatsApp for customer support saw it in action and asked if they could use it too.

The funny part is that it's open source

@garrytan But 7+-3 only works for the conscious observer, right? Otherwise we'd have a hard time walking, or talking :D

The subconscious can paint, play music, pattern map then use said map to leave the map into new territory.

Conscious sees AI as replacement. Subconscious as a new tool.

@garrytan But where’s the actual credit? And who owns what? U would have to prove ownership those are not original ideas if they already exist?????

@garrytan It's hard to understand how bad it can be at remembering stuff though.

@garrytan GBrain turns AI agents into true thinkers. Just-in-time genius over dumb context dumps the librarian that builds superintelligence.

@garrytan Big unlock. Big obstacle (gov't regulation) in its way!

@garrytan The capacity unlock is real. But latency is the actual bottleneck in practice. I built trading bots that could see a year of market data but needed fresh prices every 30 seconds. Stale data kills you faster than no data.

If retrieval is slower than task pace, you timeout.

@garrytan most agent frameworks still brute-force the wrong files into context. 200k tokens of irrelevant code is worse than 20k of the right files. selection > size

@garrytan This is peak

@garrytan but it has them in context, but it’s not deterministic recall of decisions made 6 months ago … that’s why I built Lore

https://github.com/itsthelore/requirements-as-code

@garrytan Gbrain unlocking agents for real bro

@garrytan there's no such thing as "context"

@garrytan how many things can you multitask on? i tend to top out at managing 5-6 projects at once but trying to push it..

@garrytan I'm loving how you are promoting gbrain these days 🙂

@garrytan been running gbrain in our internal agents for a bit now. it's the first retrieval layer that held up under real load. the librarian framing is right. context selection was our ceiling and this is the first thing that actually moved it.

@garrytan Maybe the next layer after context selection is reading strategy.

Once the right docs are loaded, agents still need to skim like humans: scan the structure, find the relevant sections, then read deeply where it matters.