Google's Open Knowledge Format: Just Markdown for AI Agents
Prism Labs
Community video
Home / Open Knowledge Format
OPEN KNOWLEDGE FORMAT
Open Knowledge Format (OKF), explained.
Knowledge for AI should not live inside one model, one chat or one vendor.
Open Knowledge Format is an open specification introduced by Google Cloud for representing knowledge as portable files that humans and agents can both understand.
Markdown. YAML metadata. Directories. Links. Git.
Okide is an independent product built around the open OKF specification.
THE CONTEXT PROBLEM
Agents became capable faster than knowledge became portable.
The model is only part of the system.
Useful agents also need the context around your work: definitions, decisions, schemas, policies, runbooks, research, constraints and the reasoning that usually lives across wikis, chats, documents and people's heads.
The problem is not that this knowledge does not exist.
The problem is that it rarely exists in a form that is durable, inspectable and reusable across tools.
Wiki
Chat
Docs
Prompts
App memory
People
fragmented context
open knowledge repository
THE FORMAT
A format, not another knowledge platform.
OKF represents knowledge with ordinary Markdown files and YAML frontmatter, organized in directories and connected with normal links.
That choice is important.
A human can open the files. An agent can read the files. Git can version the files. Another tool can consume the files.
The format does not require your knowledge to live behind a proprietary API.
knowledge/
├── index.md
├── company/
│ ├── index.md
│ └── strategy.md
├── product/
│ ├── index.md
│ ├── positioning.md
│ └── roadmap.md
└── research/
├── index.md
└── market.md
--- type: Concept title: Product positioning description: How we describe the product and its category. status: stable tags: [product, strategy] --- # Product positioning Okide is the IDE for knowledge repositories. See [Strategy](../company/strategy.md).
An illustrative OKF-style example. It is not a redefinition of the specification. Exact conformance lives in the canonical SPEC.md.
Built from things that already work.
Markdown
Knowledge remains readable without Okide or any specialized client.
YAML metadata
Structured signals live next to the knowledge they describe.
Directories
Knowledge can be explored progressively instead of loaded all at once.
Links
Normal Markdown links turn the directory tree into a knowledge graph.
Git
History, diffs, review and rollback come from an ecosystem that already exists.
The innovation is not inventing five new primitives. It is giving them a shared format for knowledge.
PORTABILITY
The knowledge is more valuable than the application around it.
Most knowledge tools make the interface easy and the exit difficult.
OKF takes the opposite position: the durable artifact is the repository.
A tool can help you create it, search it, validate it or serve it — but the files remain useful without that tool.
Human-readable
Open a file and read it.
Agent-readable
Give the same file to a model without translating a proprietary record.
Versionable
Review knowledge changes like code changes.
Portable
Move the repository without migrating the meaning of the repository.
OKF v0.2
Open does not mean untrusted.
Once agents can generate knowledge at machine speed, provenance and freshness stop being optional concerns.
OKF v0.2 adds structured signals that help a consumer decide what it is looking at before spending tokens reading the entire document.
01
Where did this come from?
sources can record the material a concept derives from instead of separating the claim from its origin.
02
Who produced and verified it?
generated and verified let consumers distinguish generated knowledge from independently confirmed knowledge.
03
Is this still current?
stale_after gives consumers an explicit date after which knowledge should be reconsidered.
04
Is this the version we should use?
status can distinguish knowledge such as draft, stable and deprecated concepts without deleting the history.
05
Was a result produced the approved way?
For cases where the method matters as much as the answer, OKF can carry signals that allow computation or output to be attested.
These are signals, not a closed scoring system.
type remains the only always-required concept field. The additional vocabulary is optional, and OKF continues to allow custom metadata.
EXTENSIBLE BY DESIGN
A standard should create interoperability without turning into a cage.
Different domains need different metadata.
A security repository, a research repository and a product repository should not be forced into exactly the same schema.
OKF defines shared conventions while allowing producers to add their own frontmatter fields.
OKF FIELDS
--- type: Research title: Market map status: draft
CUSTOM FIELDS
# your own domain metadata market: fintech geography: LATAM confidence: medium owner_team: strategy ---
Custom fields belong in the file with the knowledge — not in a hidden Okide database. The sample fields above are not standardized OKF fields.
FOR AGENTS
Context should be navigated, not dumped.
Loading an entire knowledge base into an agent context is expensive, noisy and often unnecessary.
OKF uses directory indexes and normal links so humans and agents can discover the repository progressively.
index.md
product/index.md
positioning.md
related concepts
Read what is relevant. Follow the next link. Keep the rest on disk.
SEE OKF IN PRACTICE
The specification ships with real examples.
The canonical Open Knowledge Format repository includes sample bundles produced around recognizable datasets and domains.
GA4 / Google Merchandise Store
E-commerce analytics knowledge around the GA4 public dataset.
Stack Overflow
A multi-concept example built around the public Stack Overflow dataset.
Acme Retail
A fictional retail knowledge repository demonstrating trust, provenance, freshness, lifecycle and other v0.2 concepts.
FROM THE SOURCE
Introduced in the open by Google Cloud.
Google Cloud introduced Open Knowledge Format in June 2026 as a vendor-neutral way to represent the context that surrounds data and systems.
The specification and its reference implementation now live in a public GoogleCloudPlatform repository.
WATCH
OKF, explained by the community.
Prefer watching to reading? These independent walkthroughs explain the format, the context problem and examples of OKF in practice.
Open Knowledge Format Explained | Google's New AI Standard
AI with Surya
Community video
WHAT OKF IS — AND ISN'T
A knowledge format plays a different role from a knowledge product.
Role
Durable artifact
OKF
RolePortable representation of knowledge.
Durable artifactMarkdown files + metadata + structure.
Wiki / documentation product
RoleInterface for people to create and browse information.
Durable artifactUsually application-owned pages or exported documents.
Vector database
RoleRetrieval infrastructure.
Durable artifactEmbeddings/indexes derived from source material.
Agent memory
RoleRuntime continuity for an agent or application.
Durable artifactUsually implementation-specific state.
Okide
RoleIDE for creating and maintaining knowledge repositories.
Durable artifactYour repository.
These things are not mutually exclusive.
An OKF repository can be the durable source of knowledge while search, retrieval, agents and interfaces operate around it.
OKF + OKIDE
A good format still deserves a good interface.
You can author OKF with a text editor.
That openness is the point.
But as a repository grows, people still need a practical way to create documents, navigate structure, edit metadata, understand trust and freshness, review history and connect the repository to agents.
That is the job of Okide.
Author
Write knowledge without spending your day hand-editing frontmatter.
Structure
Navigate the repository as knowledge, while keeping the underlying files ordinary.
Validate
Catch format and repository problems before they become agent context.
Review
See lifecycle, freshness, sources and trust signals where you work.
Version
Keep the repository backed by Git and preserve meaningful history.
Connect
Let compatible agents access the same repository through MCP.
The repository is the product you keep. Okide is the environment you use to work on it.
For agents specifically: coding agents, research agents, internal assistants, agent-generated knowledge.
Questions.
Short answers. The specification remains the authority.
What is Open Knowledge Format?
Open Knowledge Format (OKF) is an open, vendor-neutral specification introduced by Google Cloud for representing knowledge as Markdown files with structured metadata, directories and links that can be used by both humans and AI agents.
Who created Open Knowledge Format?
Open Knowledge Format was introduced by Google Cloud. Its canonical specification and reference implementation are published in the public GoogleCloudPlatform open-knowledge-format repository.
Is OKF a Google product?
OKF is an open specification published through Google Cloud and its public repository. Okide is an independent product built around the format and is not a Google product.
What files does OKF use?
The format is based on normal Markdown documents with YAML frontmatter, arranged in directories and connected with links. Consult the canonical specification for exact conformance requirements.
Does OKF require a specific AI model?
No. The format is intentionally vendor-neutral and is not tied to a particular model provider, agent framework or serving system.
Can I use OKF without Okide?
Yes. That is fundamental to the idea. An OKF repository remains ordinary files that can be read or edited with other compatible tools.
What did OKF v0.2 add?
Version 0.2 expanded the vocabulary for agentic trust, including structured signals for provenance, verification, freshness, lifecycle and attestation while keeping those additions optional.
Can OKF metadata be extended?
Yes. The specification is intentionally minimally opinionated and permits additional frontmatter fields. Consumers should not require every domain to fit into one closed metadata schema.
Is OKF a vector database or RAG framework?
No. OKF represents knowledge. Retrieval systems, vector databases and agent frameworks can consume that knowledge, but they solve a different layer of the system.
Why use Git with an OKF repository?
Because the repository is made of files, Git naturally provides history, diffs, review, branching and rollback without inventing a separate versioning mechanism for knowledge.
Where does Okide store my knowledge?
The desktop product works with the repository as files on your computer. The files remain readable independently of Okide.
How do agents use an Okide repository?
Agents can consume the same open knowledge artifacts humans maintain. Compatible agent workflows can also connect through Okide's MCP capabilities rather than requiring a second proprietary knowledge copy.
Start with an open repository.
Markdown you can read. Git history you can inspect. Knowledge your agents can use.
No account required for the desktop IDE.
Open knowledge deserves a great IDE.