Google Open Knowledge Format (OKF) — the LLM-Wiki Pattern, Standardized

Source: ai-research/google-open-knowledge-format-okf-2026-06-13.md (the OKF v0.1 SPEC.md + okf/README.md from GoogleCloudPlatform/knowledge-catalog, plus the Google Cloud announcement blog, all 2026-06-12) — surfaced via an X bookmark from @Marie_Haynes (raw/x-bookmarks-recent-digest-2026-06-13.md).

On 2026-06-12, Google Cloud published the Open Knowledge Format (OKF) — an open specification that represents knowledge as a directory of markdown files with YAML frontmatter. The announcement (by Sam McVeety and Amir Hormati) explicitly frames OKF as formalizing “the LLM-wiki pattern into a portable, interoperable format,” and cites Andrej Karpathy by name as inspiration. In other words: Google has standardized the exact pattern this vault has been running for months. OKF is vendor-neutral, requires no SDK or central authority, and is designed so any producer (human, agent, or export pipeline) can write it and any consumer (Obsidian, an LLM’s context window, a search index) can read it.

Key Takeaways

• OKF is “just markdown + files.” A Knowledge Bundle is a directory tree of markdown documents (each a Concept), each with a YAML frontmatter block and a free-form body. “If you can cat a file, you can read OKF; if you can git clone a repo, you can ship it.” No schema registry, no central authority, no required tooling.
• Exactly one required frontmatter field: type — a short, self-explanatory string (e.g. BigQuery Table, API Endpoint, Metric, Playbook, Reference). Type values are not registered centrally; consumers must tolerate unknown types gracefully. Recommended-but-optional fields: title, description, resource (canonical URI), tags, timestamp (ISO 8601). Producers may add arbitrary keys; consumers must preserve unknown keys and not reject unrecognized fields.
• Reserved filenames mirror this vault’s own structure. index.md = a directory listing for progressive disclosure (navigate one level at a time instead of loading the whole bundle); log.md = chronological update history. Every other .md is a concept. This wiki already runs _index.md per topic and an append-only log.md — OKF independently arrived at the same two load-bearing files.
• Graph-shaped, not just tree-shaped. Concepts relate via standard markdown links (richer than the parent/child the directory implies) and cite external sources — the same [[wikilink]] + **Source:** discipline this vault enforces, expressed in vanilla markdown links.
• Producer/consumer independence is the whole point. Anyone can produce OKF (humans by hand; agents on Google ADK / LangChain / custom; export pipelines from Dataplex / Unity Catalog / Collibra; scripts walking a database) and anyone can serve/consume it (static file server, Obsidian / Notion / MkDocs, an LLM loading files into context, a search index, a graph viewer). The format is the contract.
• Shipped with a reference implementation, not just a spec. An enrichment agent (Google ADK + Gemini) walks BigQuery datasets and auto-drafts OKF docs with citations and join paths (a two-pass design: a BigQuery-metadata pass + an LLM web-crawl pass with --web-max-pages and allowed-hosts caps), plus a backend-free static HTML visualizer and three browsable reference bundles (GA4 e-commerce, Stack Overflow, Bitcoin).
• v0.1 is explicitly “a starting point, not a finished standard,” designed for backward-compatible growth. Non-goals: it does not define a fixed taxonomy, prescribe storage/serving/query infra, or replace domain schemas (Avro/Protobuf/OpenAPI) — it references them.
• Karpathy’s thesis, quoted by Google: “LLMs don’t get bored, don’t forget to update a cross-reference, and can touch 15 files in one pass. The bookkeeping that causes humans to abandon personal wikis is exactly what LLMs are good at.” This is the same justification that underpins this vault and every entry in the Karpathy Pattern topic.

Why this matters for this wiki

OKF is the first vendor-backed standardization of the Karpathy LLM-wiki pattern — a stronger signal than any community implementation tracked so far (Agent Wikis, synthadoc, Stride). Where Agent Wikis is a commercial hosted take and synthadoc is a community toolchain, OKF is a portable file format — which means a wiki authored here could be consumed by a Google ADK agent, and vice-versa, with no translation layer.

This vault already substantially conforms to OKF’s structural conventions: per-directory index files, an append-only log.md, per-article YAML frontmatter, markdown-link cross-references, and a **Source:**/citation discipline. The notable difference is strictness in the other direction — this vault’s frontmatter schema (title, topic, tags, sources, created, updated, confidence, provenance, type, status) is far richer than OKF’s single required type. Because OKF explicitly allows arbitrary extra keys and requires consumers to preserve them, this vault’s articles are nearly OKF-conformant already: the main gap is that OKF reserves type as REQUIRED with concept-kind values (e.g. Playbook, Reference), whereas this vault’s type field uses pipeline-stage values (source-summary, concept, connection, file-back). The two are reconcilable.

Try It

• Read the spec, then diff it against this vault. Open okf/SPEC.md in GoogleCloudPlatform/knowledge-catalog. Map each OKF convention (index.md, log.md, frontmatter type, markdown-link relations, # Citations) onto this vault’s equivalents. The overlap is the evidence that the pattern is now a de-facto standard.
• Make a bundle OKF-consumable. To let any OKF-aware agent read this wiki, the lowest-effort move is to ensure every article keeps a type frontmatter field (already present) and that _index.md files act as directory listings (already true). No content rewrite needed.
• Convert existing markdown to OKF. The nascent chapter42/okf-convert (a Claude Code skill + standalone CLI) deterministically converts markdown files and web pages into a conformant OKF v0.1 bundle with optional LLM enrichment — a quick way to test interop without hand-editing.
• Point an agent at an OKF bundle. Clone one of Google’s reference bundles (bundles/ga4/, bundles/stackoverflow/, bundles/crypto_bitcoin/) and have a coding agent answer questions from it using only the index.md → concept traversal. This is the consumption side of the same progressive-disclosure pattern this vault uses for retrieval.
• Track adoption. Early ecosystem repos (superops-team/okf, okf-convert, okf-roam for Emacs) are day-old and tiny; the canonical artifact is Google’s repo. Watch whether OKF picks up non-Google producers — that’s the test of whether it becomes the interop layer it aspires to.

Open Questions

• Will OKF converge with AGENTS.md / skills conventions? OKF standardizes the knowledge layer; AGENTS.md and the skills ecosystem standardize the instruction layer. Whether they merge or stay parallel is unresolved.
• Does the single-required-type-field minimalism survive contact with real corpora? This vault found it needed 10 frontmatter fields + a contradiction/status workflow to stay healthy at 450+ articles. OKF’s “minimally opinionated” bet may need to grow exactly the fields this vault already added.
• Who else produces OKF? As of 2026-06-13 the only non-trivial producer is Google’s own BigQuery enrichment agent. The standard’s value is proportional to multi-vendor adoption.

Related

• Karpathy Pattern (topic index) — community + commercial implementations of the LLM-maintained-wiki pattern OKF now standardizes.
• Agent Wikis — the commercial hosted take on the same pattern; OKF is the open file-format counterpart (a bundle could feed Agent Wikis or vice-versa).
• synthadoc — community toolchain for building/maintaining LLM wikis; OKF is the interchange format such tools could emit.
• Karpathy’s LLM-Wiki Pattern — the original technique this vault implements and OKF formalizes.
• The Agent-Readable Web — OKF is the knowledge-layer instance of the broader “ship structured artifacts agents can read directly” thesis (alongside llms.txt, WebMCP, AGENTS.md).
• The 2026 Claude Code AIOS Pattern — the convergent operator pattern whose knowledge layer OKF could become the portable substrate for.
• Skills Ecosystem — the instruction-layer standardization effort running in parallel to OKF’s knowledge-layer one.