Verdict: Google’s Open Knowledge Format (OKF) is a vendor-neutral specification that formalizes how AI agents represent, share, and consume organizational knowledge. By organizing data into a standardized directory of Markdown files with YAML front-matter, OKF slashes token consumption and eliminates the navigation errors common in unmanaged "second brains" or complex multi-agent orchestration systems.
Last verified: 2026-06-26 · Status: v0.1 Specification · Best for: Developers building multi-agent systems and small businesses managing growing AI knowledge bases. Note: Volatile facts like tool availability and cloud integrations are current as of June 2026.
What is Google’s Open Knowledge Format (OKF)?
Launched on June 12, 2026, the Open Knowledge Format (OKF) is an open standard designed to solve the "fragmented knowledge" problem. Currently, most AI agents struggle to find information scattered across wikis, code comments, and database APIs. OKF turns this scattered context into a portable, version-controlled graph of "concepts."
Unlike proprietary databases, an OKF bundle is just a directory of Markdown files. If you can use Git and a text editor, you can manage an OKF knowledge base. This "LLM-wiki" pattern—originally popularized by Andrej Karpathy—is now a formal standard backed by Google Cloud.
The 3 Principles of OKF
According to the official Google Cloud specification, OKF is built on three core pillars:
- Minimally Opinionated: Every document requires only one field:
type. Much like standardized AI skills, this allows you to define your own data models while maintaining interoperability across platforms. - Producer/Consumer Independence: A human can hand-author a bundle, and an AI agent can consume it. Conversely, an agent can synthesize a bundle for a human to browse.
- Format, Not Platform: OKF does not require a specific cloud, database, or SDK. It is a contract for how files should be structured, making knowledge portable across OpenAI, Anthropic, and local models.
How OKF Works: Concepts, Folders, and index.md
OKF represents knowledge as concepts. A concept can be a database schema, a business process, a technical guide, or a product description.
1. The Folder Structure
Information is organized into folders named after the topic. For example, a folder named /billing/ would only hold concepts related to billing logic.
2. The index.md File
In every folder (including the root), an index.md file acts as a map. It provides the agent with a high-level summary of what is available in that directory. This allows for progressive disclosure: the agent reads the index first and only opens specific files if they are relevant, drastically reducing token waste.
3. YAML Front-Matter
Every concept file starts with a small YAML block:
---
name: "Refund Policy"
type: "policy"
description: "Rules for issuing customer refunds in 2026"
tags: ["finance", "customer-service"]
---
This metadata tells the agent exactly what the file holds without requiring it to "read" the entire body.
OKF vs. RAG: Why Curation Matters
While Retrieval-Augmented Generation (RAG) is useful for searching large datasets, it has a major flaw: it re-derives knowledge from raw text chunks every time you ask a question. This often leads to "AI slop"—generic, repetitive answers that miss the nuances of your business.
OKF is the "curated layer." Instead of just indexing chunks, you (or your agents) store interlinked concepts that are version-controlled and updated directly. This results in:
- Faster Retrieval: Agents navigate the index rather than scanning thousands of vectors.
- Lower Costs: Fewer tokens are spent "searching" for context.
- Higher Accuracy: Pre-structured context prevents the agent from misinterpreting raw data.
Getting Started: The Google OKF Reference Tools
Google has released several reference implementations to help teams adopt the format immediately:
| Tool | Purpose | Source |
|---|---|---|
| Enrichment Agent | Crawls BigQuery datasets and drafts OKF concept docs automatically. | Google Cloud Blog |
| HTML Visualizer | A single-file tool that turns an OKF bundle into an interactive graph view. | GitHub Repository |
| Sample Bundles | Pre-built bundles for GA4, Stack Overflow, and Bitcoin datasets. | Knowledge Catalog Repo |
What this means for you
If you are a small business owner or a developer using AI agents, OKF is your blueprint for Permanent Memory. Instead of re-explaining your guidelines to Claude or ChatGPT every morning, you can maintain an OKF bundle in a GitHub repo. This is the next evolution beyond a simple AI memory system built in Obsidian; it's a shared, scalable language for your entire agentic workforce.
FAQ
Q: Do I need a Google Cloud account to use OKF? A: No. OKF is a vendor-neutral standard. You can host your OKF bundle on GitHub, a local drive, or any cloud provider.
Q: How does OKF compare to llms.txt?
A: llms.txt is a "map" that points agents to key pages on a website. OKF is the "territory"—it holds the structured content itself in a way agents can traverse like a graph.
Q: Can I use OKF with Obsidian? A: Yes. Since OKF is just Markdown and YAML, it works perfectly with Obsidian. Many developers use the OKF folder structure to organize their Obsidian vaults for AI compatibility.
Q: Does OKF replace Vector Databases? A: Not entirely. Vector DBs (RAG) are still better for massive, unstructured datasets. OKF is for your "core" knowledge—the high-value context that must be 100% accurate.
Discussion
0 comments