Open Knowledge Format (OKF)

A vendor-neutral open specification from Google Cloud that formalizes the Karpathy LLM-wiki pattern — a directory of Markdown files with YAML frontmatter, one concept per file, cross-links as the graph — into a portable, interoperable standard for AI-agent knowledge. Published v0.1 on 2026-06-12. Source: Google Cloud Blog, https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing, 2026-06-16

What it is

OKF is a format, not a service or platform. An OKF bundle is a directory of Markdown files where each file is one concept (a table, dataset, metric, playbook, runbook, API, or anything else worth capturing), the file path is the concept's identity, and concepts link to each other with ordinary Markdown links — turning the directory into a knowledge graph richer than filesystem parent/child nesting. There is no schema registry, no central authority, no required SDK or runtime: "If you can cat a file, you can read OKF; if you can git clone a repo, you can ship it." Source: GoogleCloudPlatform/knowledge-catalog okf/SPEC.md, 2026-06; Google Cloud Blog, 2026-06-16

It targets the context-assembly problem: before an agent can answer "how do we compute weekly active users from our event stream?", it has to gather context scattered across a metadata catalog, a shared drive, code comments, a Confluence wiki, and a senior engineer's head. OKF is the curated, version-controlled layer that holds that knowledge in a form any agent can load directly. Source: TechTimes, https://www.techtimes.com/articles/318416/20260615, 2026-06-15

The spec (v0.1)

• Bundle = directory of Markdown files with YAML frontmatter; one concept per file; path = identity. Source: Google Cloud Blog, 2026-06-16
• Minimally opinionated — exactly one field is required on every concept: type. Everything else (what types exist, what other fields to include, what body sections) is left to the producer. The spec defines the interoperability surface, not the content model. Source: okf/SPEC.md, 2026-06
• Reserved structured fields: type, title, description, resource, tags, timestamp. Type values are not registered centrally; consumers MUST tolerate unknown types gracefully. Source: MarkTechPost, https://www.marktechpost.com/2026/06/16/, 2026-06-16
• Cross-links — concepts reference each other with normal Markdown links, forming the graph.
• Reserved files — optional index.md (progressive disclosure to manage the agent's context window) and log.md (change history). Source: MarkTechPost, 2026-06-16
• Producer/consumer independence — a human-authored bundle can be read by an agent; a pipeline-generated bundle can be browsed in a visualizer; an LLM-synthesized bundle can be queried by another LLM. The format is the contract; tooling at each end is swappable. Source: Google Cloud Blog, 2026-06-16

Google shipped reference tooling alongside the spec: a BigQuery enrichment agent, a static HTML visualizer, and three sample bundles, in the GoogleCloudPlatform/knowledge-catalog repo. The source repo was at d44368c15e38e7c92481c5992e4f9b5b421a801d on main, Apache-2.0, with 6,083 stars and 485 forks at the 2026-07-03 check. Source: GitHub GoogleCloudPlatform/knowledge-catalog, 2026-07-03

OKF vs RAG

OKF is complementary to, not a replacement for, RAG. RAG retrieves chunks from a large, dynamic, unstructured corpus at inference time via vector search; OKF is a curated, stable, pre-compiled knowledge layer agents read and update directly. OKF gives RAG cleaner structured source material and adds relationship context raw chunks lack. Source: TechTimes, 2026-06-15 See Context Engineering.

Lineage: Karpathy's LLM Wiki

OKF explicitly traces to Andrej Karpathy's LLM Wiki three-layer model: (1) raw material — immutable source (papers, schemas, minutes); (2) wiki body — the LLM-owned layer of cross-linked Markdown pages; (3) schema — the config that tells the agent the wiki's structure and etiquette (where AGENTS.md / CLAUDE.md live). OKF standardizes layer (2) — how the knowledge itself is expressed — while AGENTS.md/CLAUDE.md remain layer (3). Source: AI-Driven Lab, https://note.com/ai_driven/n/n8e2726b98180, 2026-06 See Karpathy LLM Wiki Setup Guide.

Relevance to kevin-wiki

This is the headline signal: an industry body just standardized the exact pattern this wiki is built on. kevin-wiki is already, in substance, an OKF-shaped bundle:

The practical implication: kevin-wiki could be made OKF-conformant and exportable with near-zero structural change (the main gap is OKF's reserved field names vs this wiki's frontmatter), making it portable to any OKF-speaking agent or visualizer. It also validates the architecture bet — the LLM-wiki pattern is becoming an interoperability standard, not a personal quirk. Worth tracking for a future "export kevin-wiki as an OKF bundle" task. Source: compiled from AGENTS.md; Google Cloud Blog, 2026-06-16

The reviewed launch image shows a concrete concept file, not only marketing copy: YAML frontmatter for a BigQuery table (type, title, description, resource, tags, timestamp), followed by a Markdown schema table and a join relationship to another concept. That visual matters because it shows OKF's interoperability contract at the file level: frontmatter gives routing metadata, while Markdown carries schema and graph relationships. Source: X/@GoogleCloudTech image, 2026-06-16; Source: local image review, 2026-07-03

Status

v0.1, published 2026-06-12, open spec in GoogleCloudPlatform/knowledge-catalog. Early — the reserved field set and conformance rules may evolve. Reference-only for now; promote to an active "conform/export" task if Kevin wants kevin-wiki interoperable with the broader OKF tooling ecosystem.

---

Timeline

• 2026-06-21 | Captured from Kevin's X bookmark of @GoogleCloudTech's OKF launch (2026-06-16). Researched primary sources (Google Cloud blog, knowledge-catalog SPEC.md) plus independent coverage (MarkTechPost, TechTimes, AI-Driven Lab). Flagged the direct correspondence between OKF v0.1 and kevin-wiki's structure. Source: X/@GoogleCloudTech, 2026-06-16; Source: Google Cloud Blog, 2026-06-16
• 2026-07-03 | Deep-reviewed the launch image and refreshed source authority: GoogleCloudPlatform/knowledge-catalog, Apache-2.0, main HEAD d44368c15e38e7c92481c5992e4f9b5b421a801d, 6,083 stars, 485 forks. The image shows the OKF concept-file shape: YAML frontmatter, Markdown schema table, and link-based join relationship. Source: raw/x-bookmarks/enriched/2067012903337664886.json; GitHub knowledge-catalog, 2026-07-03