Google’s New Release Just Fixed AI Systems
🔗 Link do vídeo: https://www.youtube.com/watch?v=k4sMSsMzX2g
🆔 ID do vídeo: k4sMSsMzX2g
📅 Publicado em: 2026-06-26T14:10:29Z
📺 Canal: AI LABS
⏱️ Duração (ISO): PT11M54S
⏱️ Duração formatada: 00:11:54
📊 Estatísticas:
– Views: 14.604
– Likes: 505
– Comentários: 20
🏷️ Tags:
Your second brain breaks down at scale, and running it like a Claude OS doesn't fix it. In this video we break down Google's Open Knowledge Format (OKF): the new standard that gives your AI agents a portable, fast-to-search knowledge base, built on the LLM Wiki pattern from Andrej Karpathy.
Community with All Resources: http://ailabspro.io
The Roundup: Our daily newsletter covering the AI stories.
Join now: https://www.theroundup.so/
If you've ever set up a second brain or tried to run your whole setup like an AI OS or Claude OS, you already know the problem. Every system is built around one person's workflow, so it's hard to share and even harder for an agent to navigate. Claude doesn't know what already exists in your knowledge base, so it searches by matching keywords, wastes tokens, and sometimes drops files in the wrong place. OKF fixes that by standardizing how knowledge is organized, the same way MCP standardized tool access, skills standardized reusable instructions, and standards like agents.md and design.md standardized agent and design context.
The problem: why your second brain stops scaling
– Why a knowledge base built around one person's workflow is hard to share and hard for an agent to navigate
– How Claude searches by matching keywords and file names, and why that falls apart on a big, deeply nested knowledge base
– Why it drops files in the wrong folder and spins up duplicates for info that already exists somewhere else
– The real cost: wasted tokens and slow retrieval that only gets worse as the base grows
What Open Knowledge Format actually is
– Why OKF is a standard, not a tool, just like MCP for tool access, skills for reusable instructions, and agents.md and design.md for context
– The LLM Wiki pattern from Andrej Karpathy it's built on, and why markdown beats RAG (the usual vector-database approach) for real AI memory
– Why RAG rebuilds the answer from scratch every time, while a markdown knowledge base lets an agent gather context as it goes
– How OKF packages your files into a portable bundle that both a human and a model can read
– Our read: why Google may be betting on structured knowledge as the next step for agentic search
How OKF works under the hood
– Concepts and bundles, and the minimalism principle: one file represents one thing
– index.md files that tell the agent what's inside each folder before it opens anything
– YAML front matter (a small label at the top of each file) that lets the agent load only what it actually needs
– Why separating the knowledge from whoever's reading it is what makes the same base work across any tool
We tested OKF on our own team's second brain
– The one we version control with Git and share through GitHub, so new teammates can pull it and get context fast
– Why the official tooling only works with Google BigQuery (Google's data warehouse), and the workaround we built
– The Markdown to OKF skill: a Claude skill that converts any folder of markdown into a valid OKF bundle, with evals to keep the conversion reliable
– The visualization tool that turns your whole knowledge base into an interactive graph view
– The claude.md change that finally got Claude Code to use the structure instead of falling back to keyword matching
What you actually get
– Lower token usage and faster retrieval, with far fewer wrong-folder mistakes
– A knowledge base both a human and a model can read, and that a new teammate can pick up instantly
– Better AI memory for your agents whether you work in Claude Code, ChatGPT, or Claude Cowork
– Where this is heading: structured knowledge as the backbone of agentic search
The same standardization problem applies everywhere
– These ideas are tool-agnostic: it's the structure that speeds up search and cuts tokens, not the model
– Whether you run Claude Code, ChatGPT, or Claude Cowork, an unorganized knowledge base wastes tokens the same way
– For anyone searching for a second brain for AI, AI agent memory, a RAG alternative, the LLM Wiki pattern, or how to build a knowledge base for AI agents: the standard is what matters, not the tool
– The Markdown to OKF skill follows the same install-and-run pattern as our other Claude skills
– Google AI looks to be betting on structured knowledge and OKF as the next step for agentic search
This is the kind of AI automation and AI OS workflow we build and test for ourselves first, so you get the real version instead of the hype. The Markdown to OKF skill, the starter packs, and our other resources live in AI Labs Pro, our community. That's where you get the skills and a place to talk with our team and other builders. The link is below.
0:00 Intro
0:32 The Problem With Second Brains
2:15 What Open Knowledge Format Is
5:12 How OKF Works
6:37 Sponsor: Mobbin
7:34 Testing It On Our Second Brain
#ai #claude #claudecode #claudeai #claudecowork #chatgpt #claudeos #google #openknowledgeformat #okf #secondbrain #aimemory #aiagents #llmwiki #knowledgebase