Every agent project eventually runs into the same problem: the model doesn't know what your organization knows. Join keys, metric definitions, deprecated tables and runbooks live in wikis, catalogs, code comments and people's heads. In June 2026 Google Cloud published a small spec for writing that knowledge down: the Open Knowledge Format (OKF). This post covers where it came from, what it did for Google, who is using it, and, because the claims got loud fast, what the performance evidence actually says.
The short version
OKF is a folder of markdown files. Each file is one concept (a table, a metric, a runbook), starts with YAML frontmatter, and links to other concepts with ordinary markdown links. The spec is about 451 lines and requires exactly one field, type (Implicator). Google's own announcement says it "formalizes the LLM-wiki pattern into a portable, interoperable format" (Google Cloud). The idea is not Google's, so the history matters.
How we got here
Retrieval first. The 2020 RAG paper gave "retrieve at question time" a name, and a vector-database boom followed from 2021 to 2023. RAG works well on large, messy corpora, but chunks can split an answer, similarity rewards text that looks relevant rather than text that is correct, and relationships between facts are inferred rather than stated.
Files as memory. In 2025, evidence started piling up that plain files could carry agent memory. Letta reported an agent using plain files scoring 74.0% on LoCoMo (versus 68.5% for Mem0), and Anthropic shipped a file-based memory tool and Agent Skills in the fall (Taskade's timeline). In February 2026, the creator of Claude Code said agentic search generally works better than RAG for it:
The spark. On April 2, 2026, Andrej Karpathy posted about using LLMs to build personal knowledge bases for research topics, as markdown wikis the model compiles and maintains:
Two days later he published an "idea file" gist called LLM Wiki. Its structure is three layers: immutable raw sources the LLM only reads, a wiki of LLM-written markdown pages (summaries, entities, concepts) that the LLM owns, and a schema document that tells the LLM how the wiki is organized (OpenKnowledge docs). People rebuilt it immediately: one write-up counts 200+ implementations and 800+ responses within two months (kfchou). Similar conventions were already common in repos too. AGENTS.md alone is reported in more than 60,000 open-source projects (Implicator).
The standard. On June 12, 2026, Sam McVeety and Amir Hormati of Google Cloud published OKF v0.1. Hormati's summary was that it is "a format, not a platform." The pitch is interoperability: a bundle written by a human, a pipeline or an LLM should be readable by any compliant tool, with no required cloud, SDK or account. v0.2 followed on July 24 with optional trust fields (who wrote a concept, who verified it, when it expires) (Taskade), and the project now lives in its own repo, GoogleCloudPlatform/open-knowledge-format, under Apache-2.0 (notes).
What it did for Google
The practical benefit is on the product side. Google updated its Knowledge Catalog to ingest OKF bundles the same day and serve them to its agents, and it shipped an enrichment agent that drafts concept files from BigQuery datasets, a static graph viewer, and three sample bundles (GA4 e-commerce, Stack Overflow, Bitcoin) (Google Cloud).
The strategic read from one analysis is worth knowing: the file format is free, while the governed layer that stores, serves and controls access to the knowledge is what Google sells, and that layer is explicitly outside the spec. Google, for its part, argues a format's value comes from how many parties speak it, not who owns it (Implicator). Google has not published internal numbers showing what OKF improved for it, so "how it helped Google" is a product-integration story, not a measured one.
Who is using it
Early signals, with the caveat that most are small:
- Launch-day reach: SEO researcher Marie Haynes posted on X that OKF "could replace Notion or Obsidian," reportedly drawing about 1M views (source).
- An independent site, No Hacks, published a hand-written OKF bundle for its own pages within a day. At launch, Google wrote all three sample bundles and both reference implementations, and the same coverage reported no outside adoption yet (Implicator).
- Tools added export: Substrate can export an OKF zip (Doug Belshaw), and Nowledge Mem exports OKF 0.2 bundles (Nowledge).
- Skeptics: an analyst quoted in the same coverage says being published is not the same as being widely used, and another commentator called a v0.1 spec from one vendor "an invitation" rather than a standard.
Did it improve performance?
This is where the hype outruns the data. Google's announcement and launch coverage include no benchmarks (Implicator). What exists is independent and mixed.
A small real-world test favors restructuring. Developer Joe Karlsson converted about 200 markdown files that teach Claude Code his homelab and tested three versions on 44 questions, three runs each (write-up):
| As-is | OKF metadata only | Restructured | |
|---|---|---|---|
| Tokens loaded at session start | 16,237 | 16,237 | 8,505 |
| Accuracy (Sonnet) | 97% | 98% | 99% |
| Cost per question (Sonnet) | $0.043 | $0.042 | $0.030 |
Adding OKF frontmatter alone changed nothing measurable. The gain came from restructuring into one concept per file behind an index, which roughly halved startup context and cut cost per question by about 30%. His own caveats: 44 questions is small, and accuracy "held" rather than improved.
A controlled RAG study found no lift. A study on SSRN compared the same source document as ordinary RAG chunks and as OKF concepts. An early result looked like a big OKF win (91.1% page hit versus 65.8% for a dense-vector baseline), but the dense encoder could read only 256 tokens while 80.9% of passages were longer, and the added OKF source did not improve retrieval or answer quality (SSRN).
Treat vendor claims carefully. One OKF generator site advertises a ~97% token reduction (45,000 to ~140 tokens per lookup) (source). That is a tool vendor's marketing number, not an independent benchmark.
My reading: OKF's measurable upside so far looks like structure and token discipline (small, indexed, linked files an agent loads on demand), not a magic accuracy boost over well-built retrieval. The format also leaves hard problems open: no access control, advisory freshness, shallow validation, and a young ecosystem.
When to use it
Curate a fact in OKF if agents rely on it on nearly every task and a wrong answer is costly: metric definitions, schemas, business rules. Leave the long tail of tickets, PDFs and logs to RAG. Start with a small pilot (a few dozen concepts, one owner, one agent use case), measure it against your current setup, and keep your source material.
Sources
- How the Open Knowledge Format can improve data sharing (Google Cloud)
- Google Open-Sources a Knowledge Format (Implicator)
- LLM Wiki Explained (Taskade)
- Can OKF make Claude Code more accurate and cheaper? (Joe Karlsson)
- Does Google's Open Knowledge Format Improve RAG? (SSRN)
- Wiki-Skills gaining traction, now with Google's OKF (kfchou)

Top comments (0)