Formaly's founder makes the markdown case for AI knowledge portability

Why it matters

Google's OKF gives AI teams a shared file format for context, and Formaly's essay shows how that standards debate reaches application startups. The companies that collect knowledge will face pressure to make their outputs portable, inspectable, and usable by agents outside their own product.

Arindam Majumder used a July 1 Formaly essay to put a founder's version of the AI knowledge-management fight on the record: context should live in files people and agents can read, rather than in retrieval systems, SDKs, and proprietary catalogs.

Majumder is coming at the issue from an unusual place. Formaly is an AI-powered form product, with a chat-first interface for collecting feedback and an analysis layer for extracting themes, sentiment, and key quotes from responses. In his February 9 introduction to the product, Majumder framed it as an AI-powered form builder that turns surveys into conversations. Introducing Formaly

That background matters because the July 1 post is not a data-catalog vendor's standards brief. It is a small AI product founder arguing that the same portability question now applies to the raw material every AI company wants to own: the user's knowledge.

The format wall Majumder is attacking

Formaly's essay gives retrieval-augmented generation its due. RAG helped when model context windows were smaller and every query needed a carefully selected slice of the knowledge base. The problem Majumder identifies is what happened around that practical fix: documents became chunks, chunks became embeddings, relationships became graph edges, and internal knowledge moved behind services that only the chosen tool could interpret.

His phrase for that is a "format wall." It is a useful frame because the lock-in here is quieter than a contract clause. A team may still own the documents, the notes, the schemas, and the runbooks. Once those materials are transformed into a system-specific retrieval stack, however, the working version of the knowledge lives behind the machinery that indexed it.

The correction Majumder argues for is deliberately plain: markdown files with enough structure to be useful and little enough structure to remain portable. He points to the pattern already visible in day-to-day agent work: CLAUDE.md, AGENTS.md, DESIGN.md, MEMORY.md, Obsidian vaults, linked notes, and project instructions that coding agents read directly. Those files are not sophisticated infrastructure. They work because a human can edit them, a model can read them, and another tool does not need permission to consume them.

For Formaly, the argument has commercial stakes. The company says its product can generate forms from prompts, collect responses through chat, embed forms on websites, and summarize response data with AI. If the output of that process becomes knowledge about churn, onboarding confusion, pricing objections, or product gaps, customers will eventually ask whether those insights can move with them. A dashboard alone is a weak answer in a market where agents increasingly operate across tools.

Google gave the pattern a name

The immediate trigger for Majumder's essay is Google's Open Knowledge Format. Google Cloud introduced OKF in June 2026, describing it as an open specification that formalizes the LLM-wiki pattern into a portable, interoperable format.

OKF v0.1 is intentionally small. A bundle is a directory of markdown files. Each file represents a concept, such as a dataset, table, metric, runbook, playbook, or API. Each file can carry YAML frontmatter for queryable fields including type, title, description, resource, tags, and timestamp. Google says type is the only required field. Relationships are ordinary markdown links, which lets a folder behave like a graph without requiring a graph database.

Google also published an OKF repository on GitHub. The repository's framing is consistent with the blog post: the format is the contribution, with reference tooling to make the specification tangible.

There is a clear incentive behind Google's move. If OKF becomes a shared convention, Google can make BigQuery and Knowledge Catalog easier for agents to consume while presenting the underlying files as portable artifacts. That does not make OKF less useful. It does mean adoption will depend on whether other producers and consumers treat it as a common format rather than a Google-shaped export path.

Majumder's essay reads OKF as important because it takes an informal behavior and turns it into a standard. That is the practical value of boring formats. Markdown plus frontmatter is not a breakthrough by itself. Agreement around the same markdown conventions can change how knowledge moves between tools.

Karpathy's wiki pattern moved first

Majumder also builds the essay around Andrej Karpathy's LLM Wiki, a plain-files pattern for letting language models maintain knowledge bases. Karpathy's setup has three pieces: a sources/ directory for immutable raw material, a wiki/ directory for model-generated markdown pages, and an instruction file such as CLAUDE.md or AGENTS.md that tells the agent how to maintain the structure.

Karpathy's argument is that the maintenance work that killed many personal wikis is exactly where language models are useful. His gist puts it bluntly: "LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass." That line is doing real work in Majumder's essay. It shifts the wiki from a human discipline problem into an agent workflow problem.

Karpathy carries weight in this conversation because his work tends to travel quickly through AI-builder circles. When a Karpathy workflow pattern gets picked up by Google Cloud and then cited by founders building AI products outside the infrastructure layer, it is no longer just a personal productivity note.

The bar Formaly is setting for itself

The strongest part of Majumder's post comes near the end, when he applies the OKF lesson back to Formaly. He argues that a survey is a knowledge-collection tool. In this framing, a response, summary, or synthesis across interviews should be readable and portable, because it represents understanding rather than a row in a spreadsheet.

That is a higher bar than most AI application startups set publicly. Many AI products promise to extract insight from messy inputs, then keep the resulting analysis inside their own interface. Formaly is arguing, at least at the level of product philosophy, that insight should remain useful after it leaves the system that produced it.

The July 1 essay outlines a thesis rather than specific export formats or implementation details that would make Formaly's own outputs portable in the way Majumder praises. It still creates an accountability point for Formaly: if knowledge should not be gated, customer feedback generated inside Formaly should eventually be easy to inspect, version, export, and feed into the next agent workflow.

That is where the founder story becomes more interesting than the standards story. Majumder built Formaly around the belief that static forms lose nuance and that AI can help teams ask better follow-up questions. The OKF essay extends that belief downstream. Collecting richer answers is only half the job. The second half is making sure the meaning extracted from those answers does not get trapped in the product that collected them.

For founders building AI applications, that is the trade to watch. Proprietary context stores can make a product feel sticky in the short term. Portable knowledge makes the product earn retention through better collection, better synthesis, and better workflows. Majumder is betting that the second path will age better.