Google published the Open Knowledge Format (OKF) late last week, a specification for representing the context, metadata, and curated knowledge that AI agents need to operate effectively. The announcement, posted to the Google Cloud blog, positions OKF as a vendor-neutral alternative to the proprietary catalog formats that have made agent knowledge fragmented and non-portable across tools and organizations.
The problem OKF addresses is real. Every team building agents today maintains its own bespoke knowledge layer: table schemas, metric definitions, runbooks, API deprecation notices. These fragments live in metadata catalogs, shared drives, code comments, and a few senior engineers’ heads. When an agent needs that knowledge, it reassembles from incompatible sources. Every agent builder solves the same context-assembly problem from scratch.
The LLM-wiki pattern that OKF formalizes is not new. Andrej Karpathy described the core insight in a public gist: LLMs are well-suited to maintaining a shared markdown library because they do not get bored, do not forget cross-references, and can update many files in a single pass. Teams have been running variations of this pattern under different names. AGENTS.md and CLAUDE.md convention files, Obsidian vaults wired to coding agents, and “metadata as code” repositories inside data teams all share the same shape: markdown, frontmatter, and cross-links. OKF adds the one thing missing: an agreed-upon set of conventions so that a knowledge bundle produced by one team or tool can be consumed by another without a translation layer.
The spec itself is deliberately minimal. An OKF bundle is a directory of markdown files, each representing one concept. Every concept carries YAML frontmatter with six queryable fields: type, title, description, resource, tags, and timestamp. The only required field is type. Concepts link to each other with standard markdown links. The full v0.1 specification, including conformance rules and reserved filenames, fits on a single page. No runtime, no SDK, no proprietary account required.
Google is shipping reference implementations alongside the spec: an enrichment agent that walks a BigQuery dataset and drafts an OKF document for every table and view, and a static HTML visualizer that renders any OKF bundle as an interactive graph without a backend. The reference agent is built on BigQuery, which is Google’s own product. That is worth noting.
The structural skepticism here is straightforward. Google is the author of a “vendor-neutral” standard whose reference implementation is seeded from BigQuery and whose first deep integration is Google Cloud’s Knowledge Catalog. A format whose adoption is led by one hyperscaler tends to embed that hyperscaler’s assumptions about what knowledge looks like and where it comes from. The format is genuinely open and requires no proprietary tooling to use. Whether the broader ecosystem, including teams running on AWS or on-premise, adopts it enough to make OKF a real interchange format rather than a Google-specific convention is the question the announcement does not answer.
The release announcement does not include any adoption commitments from outside Google, nor does it cite independent organizations piloting OKF before publication.
For teams building knowledge bases for agents today, the core idea is worth evaluating on its own merits, independent of Google’s participation. If your team is already maintaining markdown-based context files for agents, writing a thin OKF conformance layer on top adds almost nothing: a type field in the frontmatter and consistent link conventions. The cost of compatibility is low, and a portable format becomes more valuable as the agent ecosystem adds more consumers.
The format is on GitHub. If your team ships an agent knowledge base in the next quarter, spending an afternoon testing OKF conformance costs little and keeps the option of interoperability open.
Google Cloud blog, published June 13, 2026.
