3 posts tagged with "knowledge-graph"

On 2026-06-12 Google Cloud published the Open Knowledge Format: a directory of markdown files with YAML frontmatter, cross-linked into a graph, written and maintained by AI agents, with a static visualizer and a one-page spec. If that description gives you déjà vu reading this blog, it should — it's the shape StrayMark has shipped for months. We arrived there from our own ideas; OKF arrived crediting the Karpathy gist the whole "LLM-wiki pattern" traces back to. Two roads, same primitives. The honest reaction isn't defensiveness — someone with Google's reach just validated the bet. But validation of the shape is not agreement about its purpose, and that's where the only interesting question lives — the one their spec answers in a single sentence.

The walking skeleton shipped fast and looked great: a force-directed graph of every StrayMark document, colored by type, rebuilding in the browser within a second of saving a file. Then we pointed it at a real corpus — Sentinel, 395 references — and 330 of them dangled. Not because the documents were broken. Because they referenced each other the way humans write, not the way a naive graph builder matches. The render was a weekend. Making the edges land took the rest of Loom M1's follow-ups, and it changed what a "broken link" even means.

StrayMark documents have always formed a graph — every related, supersedes, originating_ailogs link is an edge. The CLI built that graph internally for straymark audit, and a human could read it one document at a time in the explore TUI. Then a second consumer showed up — Loom, an experimental browser view of the whole corpus — and asked to parse the same documents. The honest answer to "can it reuse the CLI's parser?" was no, because the parser wasn't a library; it was buried in cli/src/document.rs. The fix shipped as cli-3.23.1 with zero user-facing behavior change. It was also the most important release of the month.