Google's OKF Makes AI Agent Knowledge Just Markdown

Google Cloud has introduced the Open Knowledge Format (OKF), a vendor-neutral open specification for how AI agents get at a company's curated knowledge. In version 0.1, that knowledge is just a directory of markdown files with YAML frontmatter — each file is a 'concept', the only required field is type, and files reference each other with ordinary markdown links to form a knowledge graph. No proprietary account, no SDK, no cloud lock-in: bundles render on GitHub, ship as tarballs, and mount on any filesystem.

The problem it targets is mundane but expensive. Most organizations scatter the context an agent needs — data definitions, runbooks, wiki pages, code comments — across systems that don't talk to each other, so every team building an agent re-solves the same 'how do I assemble context' problem from scratch. OKF tries to make that knowledge portable and interoperable instead of trapping it in one platform.

How it differs from RAG

If you've followed AI tooling, you know retrieval-augmented generation (RAG): you dump documents into a database, and at query time the system grabs raw chunks and stuffs them into the prompt. OKF takes a different stance. Instead of raw chunks pulled on the fly, it stores curated, cross-linked concepts that an agent reads — and can update — directly. Google leans on three design principles: minimally opinionated (one mandatory field), producer/consumer independence (humans and agents both read and write the same files), and 'format, not platform.' To get people started, it shipped reference tools: a BigQuery enrichment agent, a static HTML visualizer, and three sample bundles.

Why this matters to you

Even if you'll never write a YAML header, this is a quiet but important bet on how AI systems remember things. Markdown plus version control means your company's knowledge becomes diff-able, reviewable and movable — the same way code already is — instead of rotting inside a Notion space nobody trusts. And because it's an open format rather than a Google product, the knowledge you put in isn't hostage to one vendor's roadmap or pricing.

Having spent a lot of time feeding context to coding agents, the part I find genuinely useful is that humans and agents read the exact same files. There's no hidden 'AI layer' to keep in sync — the thing the agent sees is the thing you can open and fix in a text editor. Whether OKF specifically wins or some rival spec does, that principle is the right one to bet on.