Open Knowledge Format: the open memory of AI agents

In June 2026, Google Cloud published the Open Knowledge Format, an open standard that represents in Markdown files the internal knowledge that AI agents need. Its promise fits in three words: a format, not a platform. For a sovereign infrastructure, this portability changes how agent memory is kept and shared.

By QDNA · Published 17 July 2026 · 8 min read

The agent landscape is fragmented today

A useful AI agent does not reason about the world in general. It reasons about the organisation's internal context: the schema of a table, the business meaning of a metric, the recovery procedure for an incident, the join path between two systems, the deprecation note for an old interface. This knowledge already exists, but it is scattered: metadata catalogues with proprietary APIs, wikis, code comments, notebook cells, and unwritten knowledge in engineers' heads.

The result is systematic waste. Every team building an agent reassembles context from scratch. Every catalogue vendor reinvents its own data model. Knowledge stays trapped inside the tool that produced it. OKF starts from a simple observation: the answer is not a new knowledge service, but a common format.

What the format standardises

OKF is deliberately minimal. Knowledge takes the form of a folder of Markdown files, one concept per file, each topped with a YAML front matter block. The format requires exactly one thing from each file: a type field, for example "table", "metric" or "procedure". The other fields, title, description, resource, tags and timestamp, are standardised but optional.

Three conventions complete the set. Files link to one another through plain Markdown links, which draws a concept graph richer than a simple tree. An optional index.md file serves as a catalogue for progressive reading. An optional log.md file keeps a chronological record of operations. The full version 0.1 specification fits on a single page, and the whole thing travels like an archive, lives in a Git repository, and reads in any editor.

Why an open format serves sovereignty

A format with no owner has properties that no service can offer. Knowledge becomes portable: it survives a change of tool, vendor or organisation. It gets versioned alongside the code it describes, with the same history and the same reviews. The same file is read by a human and traversed by an agent, with no translation layer. Nobody needs a proprietary account or an SDK to produce it, read it or serve it.

These properties map exactly onto the requirements of a sovereign infrastructure. Knowledge stays within the certified perimeter, on the same footing as data and models. OKF's reference viewer illustrates the principle: a self-contained HTML file turns a folder into an interactive graph, with no server and no data ever leaving the page. Portability stops being a contractual promise and becomes a property of the format itself.

What OKF does not do

Clarity requires a boundary. OKF is a representation and exchange format, not a search engine or a reasoning system. It carries no formal ontology and performs no retrieval. Finding the right fragment at the right moment remains the job of a dedicated layer: document vectorisation, result reranking, traversal of the relationship graph. OKF sits upstream, as the structured source that this layer indexes.

This division of labour is a strength, not a shortcoming. The format fixes the surface of interoperability and leaves the content model to the producer, retrieval to the consumer. A producer writes its files, a consumer indexes them, and the two evolve without needing to coordinate.

Agent memory, one building block of the platform

A complete agentic platform is not limited to models and orchestration. It adds a memory layer: the knowledge base that agents read, enrich and share over time. QDNA operates this layer within the organisation's own infrastructure, alongside multi-agent orchestration and the model gateway.

An open format such as OKF strengthens this approach. The memory of a sovereign platform benefits from being expressed in a format with no owner: it stays exportable, auditable and interoperable, without depending on a single vendor. The organisation keeps control of its knowledge, moves it if it changes tools, and shares it between agents with no translation. It is the same logic that guides the rest of the platform: interchangeable open models, a single gateway, and knowledge held in an open format.

An organisation's knowledge is worth more in an open format than in a proprietary database: portable, versionable and readable by humans and agents alike, it stays under the control of whoever produces it.

Adopting the format without waiting

Version 0.1 is a starting point, not a fixed standard. Nothing prevents starting small. A team exports a first scope, a few tables, a few metrics, a handful of procedures, into a folder of Markdown files with typed front matter. It versions this alongside its code, links it with plain links, and serves it to its agents. The knowledge base grows as agents read and add to it.

The value of a format lies in how many parties speak it, not in who owns it. For an organisation building a sovereign AI platform, adopting an open knowledge format means preparing tomorrow's interoperability without giving up today's control.