Google Cloud has proposed the Open Knowledge Format, or OKF, a specification for handing AI agents curated context about an organization. The version 0.1 spec is deliberately plain: knowledge is a directory of Markdown files, each carrying a small YAML frontmatter header whose only required field is type, with optional fields like title, description, tags, and timestamp. Concepts link to each other with ordinary Markdown links, and those links form a knowledge graph the agent can read and, importantly, update. The problem it targets is what Google calls the fragmented context problem: the runbooks, wikis, metadata catalogs, and code comments that hold a company's real knowledge, all siloed in systems that do not talk to each other.
The framing is pitched directly against retrieval-augmented generation. RAG answers a question by re-deriving knowledge from raw chunks at query time; OKF instead stores curated, cross-linked concepts that a model consumes directly, no translation step in between. And it asks for almost no infrastructure: no vector database, no backend, no install. A bundle is just files, so it works in version control, renders on GitHub, and travels as a tarball, and Google shows a minimal consumer written with nothing but Python's standard pathlib, re, and yaml.
The word Google leans on is neutrality. It calls OKF a format, not a platform, tied to no cloud, no database, no model provider, and no agent framework, and says it will never require a proprietary account to read, write, or serve. To seed adoption it shipped reference tooling, a BigQuery enrichment agent, a static HTML visualizer, and three sample bundles, and it is explicit that OKF is not a new idea so much as a standardization of conventions people already improvise: AGENTS.md and CLAUDE.md files, Obsidian vaults, and the LLM-wiki pattern Andrej Karpathy sketched in an April gist.
The honest question is adoption, because a format is only worth as much as the number of tools that speak it, and so far the support is Google's own. But the instinct is aligned with where a lot of practitioners already sit, and with the rest of this week's interoperability moves, from meta-harnesses that sit above agent runtimes to the push for shared agent protocols. Curated, human-readable knowledge that an agent edits in place is a more legible and more portable thing than an opaque vector store, and a vendor-neutral spec, if anyone else adopts it, is how that stops being every team's bespoke folder of Markdown and starts being something agents can expect to find. Disclosure: CLAUDE.md, cited here as prior art, is a context convention for Claude, the AI model that wrote this, made by Anthropic.