What this covers

This eight-minute explainer from Devsplainers is narrated by a single unidentified speaker. It walks through what Google Cloud published on June 12th as the Open Knowledge Format, and where the idea came from. The setup is the two-year consensus around RAG: chunk documents, turn the pieces into embeddings, and store them in a vector database. The speaker's complaint is that every query starts from zero, so the model re-derives the same connections each time. The alternative is Andrej Karpathy's LLM Wiki, in which a model builds a folder of interlinked plain-text files once and humans only supply material and ask questions. Karpathy's framing is that Obsidian is the IDE, the LLM is the programmer, and the wiki is the code base.

The middle of the video covers the spec itself. A bundle is a folder, each file holds one concept, paths serve as names, links form a graph, and two special files list contents and log changes. The one hard rule is that every file declares its type. The speaker gives three reasons this beats RAG: the work is done up front, a short table of contents lets the model skip thousands of files, and plain text in Git can be diffed, reviewed, zipped, and run offline.

The sharper material is the list of catches. The type field is free-form, so the format standardizes the container and not the meaning. Nothing in it keeps knowledge current, and Google dropped Karpathy's maintenance instructions. Models also botch markdown, and the spec's response is to tell readers to forgive broken links and unparseable files. The speaker also notes that the format came from the BigQuery team rather than Google's AI lab, with samples and tooling tied to BigQuery and Gemini. The video separates OKF from MCP, calling it cargo rather than pipe, and from SEO. It closes doubtful that the standard will stick.

Sharpest takeaway

The AI industry's two-year investment in vector databases for machine memory was rendered obsolete by a simpler approach—plain text files organized in folders—which Google has now standardized as the Open Knowledge Format, though the format's success depends on invisible organizational skill that no specification can provide.

  • RAG (Retrieval-Augmented Generation) with vector databases was the industry standard but required re-deriving connections on every query, while text-file folders build knowledge once and reuse it
  • Google's Open Knowledge Format standardizes the container but not the meaning, leaving the critical moat—organizational structure and maintenance discipline—unspecified and invisible
  • The format's permissive design and lack of update mechanisms mean it will go stale on shared teams, and language models are poor at maintaining clean markdown at scale

The claims · ranked22 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

RAG systems require that each query restart from zero, with the model receiving a fresh pile of disconnected document snippets and having to rework the same logical connections repeatedly across multiple queries rather than building persistent knowledge once.

causalhigh valueestablishednovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“It works, but it never remembers. Every query starts from zero. The model gets a fresh pile of disconnected snippets and has to work out the same connections it worked out an hour ago, and the hour before that.”

0.69

On June 12th, Google Cloud published the Open Knowledge Format, a formal specification that standardizes Karpathy's loose community idea of using plain text files for knowledge organization.

factualhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“On June 12th, Google Cloud took that loose community idea and published it as a formal spec, the Open Knowledge Format.”

0.64

For two years, the entire AI industry agreed on using vector databases and chunked documents as the standard method for giving large language models memory, converting document fragments into numerical embeddings stored in special databases.

factualhigh valueestablishednovelty 1/4durability 2/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“For 2 years, the entire AI industry agreed on how to give a model memory. You chopped your documents into thousands of fragments, turned each fragment into a long list of numbers, and paid to store them in a special database.”

0.61

The Open Knowledge Format specification is intentionally minimal, defining a bundle as a folder, with each file representing one concept, file paths serving as names, links between files forming a graph, two special file names (one listing contents and one logging changes), and one hard rule: every file must declare its type in one field.

factualhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The spec is almost comically small. A bundle is a folder. Every file is one concept, a table, a metric, a playbook, whatever you've got. The file's path is its name. Links between files form a graph. There are two special file names, one that lists what's in a folder and one that logs changes. And there's exactly one hard rule. Every file has to say what type of thing it is in one field.”

0.60

The deepest catch is that the Open Knowledge Format standardizes the container, not the meaning—the one required field is a free-form label, so one team writes 'BigQuery table' while another writes 'table' and another writes 'relational asset,' and all are valid, but each speaks a different language and agreeing on what's inside is still the team's responsibility.

causalhigh valuespeaker onlynovelty 3/4durability 4/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The deepest catch is the last one. The format standardizes the container, not the meaning. The one required field is a free-form label. Your team writes BigQuery table. Mine writes table. Someone else writes relational asset. They're all valid, but each is speaking a different language. You can ship the box anywhere. Agreeing on what's inside is still on you.”

0.60

An agent is basically just a folder of markdown files, and while anyone can write markdown, the skill is in how the folder is organized—what's locked versus what the AI can rewrite, what stops it drifting over a long run—and this organizational moat is invisible and cannot be discerned by reading the files.

causalhigh valuespeaker onlynovelty 3/4durability 4/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“An agent is basically just a folder of markdown files. Anyone can write markdown. The skill is in how the folder is organized. What's locked versus what the AI can rewrite. What stops it drifting over a long run. The moat is invisible. Two folders can look identical. One holds up in production while the other slowly rots. You can't tell which is which by reading the files. No format can hand you that.”

0.57

In spring (implied to be 2024), Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, proposed an alternative approach called the LLM Wiki using plain text files organized in folders rather than vector databases, and this approach worked better than the industry standard.

factualhigh valuespeaker onlynovelty 3/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“Then, this spring, one of the most respected names in AI looked at all of it and said, 'Just use a folder of text files.' And it worked better.”

0.56

The underlying idea that plain text folders are a better approach to machine memory than vector databases has already won, regardless of what happens to the Open Knowledge Format specifically, because the most overfunded field in tech spent two years and a fortune convincing itself that giving machines memory needed exotic infrastructure when a tidy folder of text files did the job better.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The idea underneath it has already won though. The most overfunded field in tech spent two years and a fortune convincing itself that giving a machine memory needed exotic infrastructure. Then a tidy folder of text files did the job better. Whatever happens to the format, that part isn't going back in the box.”

0.52

The Open Knowledge Format specification orders any tool reading a bundle to forgive almost everything—unknown fields, broken links, and files that cannot be parsed—making it an enterprise standard from Google whose defining feature is how little it demands and how much it lets you break.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The spec also orders any tool reading a bundle to forgive almost everything. Unknown fields, broken links, even files it can't parse. An enterprise standard from Google whose defining feature is how little it demands and how much it lets you break.”

0.52

Google dropped Karpathy's instructions for how the AI maintains the wiki from the Open Knowledge Format specification, keeping the folder structure but removing the mechanism that keeps the knowledge current.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“One thing Google dropped though, Karpathy's instructions for how the AI maintains the wiki. They kept the folder and left out the part that keeps it alive.”

0.52

The primary catch with the Open Knowledge Format is that the specification has no mechanism to keep knowledge current; while it includes a field for a timestamp, a field is not a process, and nothing in the format updates itself, so on a shared team folder it will go stale in a month because nobody volunteers to tend it, and agents start answering from expired knowledge.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The pitch is that the AI does the bookkeeping humans always abandon, but the spec has no mechanism to keep anything current. There's a field for a timestamp, fine. A field is not a process. Nothing in the format updates itself. This works beautifully when one person owns one folder. On a shared team folder, it goes stale in a month. Nobody volunteers to tend it, and the agent starts answering from knowledge that expired back in spring.”

0.52

The Open Knowledge Format did not come from Google's AI lab but from Google's BigQuery team, every sample data set ships on BigQuery, the reference tool that writes the bundles runs on Gemini, and the easiest place to pour a finished bundle is Google's own knowledge product that they just renamed for exactly this moment.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“OKF didn't come out of Google's AI lab. It came from the BigQuery team. Every sample data set ships on BigQuery. The reference tool that writes the bundles runs on Gemini. And the easiest place to pour a finished bundle is Google's own knowledge product. The one they just renamed for exactly this moment.”

0.52

The pitch of the Open Knowledge Format is that the AI does the bookkeeping that humans always abandon, which rests on an assumption about human behavior that the specification itself provides no mechanism to enforce.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The pitch is that the AI does the bookkeeping humans always abandon, but the spec has no mechanism to keep anything current.”

0.50

The text-file folder approach wins over RAG for three reasons: first, RAG does its thinking at query time while the wiki does it once upfront, paying the computational cost once and then just reading finished answers; second, a model can hold only so much in its head at once, so each folder carries a short table of contents allowing the model to read the contents, pick the file it needs, and skip the other thousands without choking; and third, it's plain text that lives in Git like code, so it can be diffed, reviewed in pull requests, zipped, and run offline on a laptop without needing a database, server, or API key.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“The folder wins and it wins for three reasons. First, when the work happens. RAG does its thinking at question time. The wiki does it once up front when the bundle is built, connecting concepts, flagging contradictions, writing the summaries. Pay that cost a single time, then just read the finished answer. Second, skill. A model can only hold so much in its head at once, and a big company has thousands of these files. Each folder carries a short table of contents. The model reads that first, picks the one file it needs, and skips the other 9,000. It never chokes on the whole library. Third, it's only text. It lives in Git exactly like code. You can diff it, review it in a pull request, zip the whole thing and hand it to a model running offline on a laptop. You don't need a database, a server, or an API key to read it. If you can open a file, you're good.”

0.50

The LLM Wiki inverts RAG by building knowledge once upfront into a folder of interconnected plain text files rather than re-deriving everything at query time, creating a living encyclopedia that models can read the way developers read a code base.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“it flips RAG on its head. Instead of the model re-deriving everything the moment you ask, you build the knowledge up once into a folder of plain text files that link to each other. A living encyclopedia the model can read the way a developer reads a code base.”

0.49

In the LLM Wiki model, the AI does the work of summarizing, cross-referencing, and filing information; humans bring new material and ask questions; and the folder of files is owned by the human while the model is the worker maintaining it.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“People hear a folder of notes and assume they're the one writing the notes. Wrong way around. You don't write the wiki, the AI does. You bring it new material and ask good questions. It handles the summarizing, the cross-referencing, the filing, all the upkeep nobody ever keeps up with.”

0.48

Google's solution to the problem of language models making errors while maintaining markdown files was to change the specification to order every reader to forgive the mess, which is damage control with a nicer name.

causalhigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“And how did Google fix the messy librarian problem? They didn't. They changed the spec to order every reader to forgive the mess. That permissive rule is damage control with a nicer name.”

0.47

The second catch is that the entire approach rests on the assumption that language models are tireless and accurate librarians, but in practice, language models are bad at writing clean markdown at scale—they botch the formatting, mangle headers, and invent links to files that were never created.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“Catch two is funnier. The whole idea rests on AI being a tireless, accurate librarian. In practice, language models are bad at writing clean markdown at scale. They botch the formatting, mangle headers, invent links to files that were never created.”

0.47

Karpathy's framing for the LLM Wiki model is: Obsidian is the IDE, the LLM is the programmer, and the wiki is the code base—with the folder as the part humans own and the model as the worker who maintains it.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“Karpathy's own line for it, Obsidian is the IDE, the LLM is the programmer, the wiki is the code base. The folder is the part you own, the model is the worker who maintains it.”

0.43

The Open Knowledge Format is not competing with MCP (Model Context Protocol); MCP is the pipe that moves data around live while the Open Knowledge Format is the cargo that moves through it.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“This isn't competing with MCP. MCP is the pipe that moves data around live. This is the cargo that moves through it.”

0.39

The day the Open Knowledge Format launched, almost nobody outside Google was using it, and a standard with one user is just a suggestion, so it could easily be the next Google project added to their graveyard.

forecasthigh valuespeaker onlynovelty 1/4durability 2/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“Will it stick? Hard to say. The day it launched, almost nobody outside Google was using it. And a standard with one user is just a suggestion. It could easily be the next Google project added to their graveyard.”

0.25

The Open Knowledge Format is not an SEO trick; nothing in it helps a search engine find you; it is private knowledge for your own agents.

factualspeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — Google's OKF: Why a Folder Beats the Vector Database [P_E29-87THI]

“and it's not an SEO trick. Nothing in it helps a search engine find you. It's private knowledge for your own agents.”