---
title: The Memory Layer & the Open Knowledge Format (OKF)
summary: These islands are written in OKF — an open, vendor-neutral standard of plain markdown files with YAML frontmatter. Thalatta uses it so the same knowledge is readable by a human and an agent, lives in version control, and belongs to no platform. Context is assembled and handed to the agent up front, not fished for mid-answer.
tags: [okf, open-knowledge-format, memory-layer, transparency, format, reference]
updated: 2026-07-14
source: https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing, about-okf.md (gated bundle)
---

# The memory layer, and the format it's written in

Every file in this bundle — this one included — is plain markdown with a few lines of YAML at
the top. The format is **OKF, the Open Knowledge Format** (plain names: an *LLM wiki*, a
*knowledge base as files*). It's the memory-layer half of how a [glass agent](glass-agent.md)'s
context is assembled.

## What OKF is

OKF is an open standard for the curated knowledge AI systems need, published by **Google Cloud's
Data Cloud team** (v0.1, June 2026). Google Cloud describes it as *"a vendor-neutral, agent- and
human-friendly standard for representing the metadata, context, and curated knowledge that modern
AI systems need"* — **their definition, attributed.** It formalizes a pattern Andrej Karpathy
called the *"LLM wiki"*: knowledge written once, in files, that both people and models can read.

The whole idea is deliberately small:

- **Just markdown.** Readable in any editor, renders on GitHub, needs no special tool.
- **Just files.** A directory you can ship as a folder, host in a git repo, or mount on a
  filesystem — no database, no runtime, no SDK, no account.
- **Just YAML frontmatter.** A few structured fields at the top of each file make it
  self-describing and machine-navigable; the prose below is for humans.
- **A knowledge graph from plain links.** Files reference each other with ordinary markdown
  links, so a bundle is a small connected wiki, not a pile of documents.

## Why Thalatta uses it

OKF is the format version of the thing Thalatta sells: **appropriate, legible context — in the
open.**

- **One file, two readers.** The exact text the agent works from is the exact text a person can
  read. No "what did the model actually see" gap.
- **Governed like code.** It lives in a git repo, so every change to what an agent says about
  Thalatta is reviewed, dated, and reversible.
- **No platform to buy.** Because it's files, not a product, the knowledge is portable and
  outlives any one vendor's tool — the same posture [Thalatta](about-thalatta.md) takes with
  every system it builds.

## Given, not searched

A glass agent never starts from zero. Before it answers, a broker assembles its context — the
caller's own record plus the right slice of the memory layer — so it begins already knowing what
the company knows. The usual alternative, handing the agent a search tool and letting it
*retrieve* chunks reactively mid-answer (traditional RAG), makes the agent fish: slow, costly,
and repeated on every run. Assembled context flips it: gather what the job needs and **hand it
to the agent before it answers**. Deterministic, auditable, and cheap, because the agent spends
its budget on the work instead of the search. (This isn't "never retrieve" — when the library is
large you still select a subset, but *up front, from a governed index you can read*.)

## A curated subset, on purpose — and tiers of depth

No bundle should hold everything a company knows. This public one deliberately doesn't: it's the
**Surface** — curated for anyone, human or AI, who sails past. A fuller briefing (the **Harbor**)
is what the live glass agent actually reads, open to signed-in visitors at **thalatta.ai/okf** —
the same files the assistant reads, which is the point of glass. And some slices are gated
deeper still, injected server-side only for people Jeff has personally flagged — no amount of
asking can summon them into a session that isn't entitled to them. Same memory layer, tiered
boundaries: one governed by curation, one enforced in the data, outside the agent's reach.

"Appropriate context, not maximum context" isn't a slogan here; it's literally how the layers
are built.