Publishing Likewise

2026-05-05

Likewise is a draft protocol for decentralized personal knowledge graphs. It’s at getlikewise.ai/spec, it’s a v0.1, and I’m publishing it tonight for review.

It specifies a wire-level format for the things personal-AI systems infer about you. Not the photos and calendar entries themselves, but the derived claims that sit on top of them: that you go to the same coffee shop on Tuesdays, that this is your child, that you and Sarah are close. Today those derived claims live in the platform that generated them, you cannot read them, and they vanish with your account. Likewise puts them on an append-only log the user owns, with capability-based delegation when they want to share a slice, and a refutation cascade when the system gets a fact about them wrong.

Twenty-five years of consumer software has agreed on the same arrangement: the party providing the service keeps the record of the user. Cookies, free accounts, loyalty cards, the personalized feed: implementation details on top of the same underlying contract. The work and the record of the work both belonged to the platform. That arrangement persisted because the record was inert raw material. A click stream could power ad targeting and ranking, but it didn’t model who you were.

What changed is that those same logs are now the training data and prompt context for systems that can describe you to yourself with uncomfortable accuracy. The economic value of being the party that holds the record has gone up by an order of magnitude, and so has the asymmetry between you and that party. Personal AI is on its way; whatever your phone is going to do for you in the next few years, the interesting layer is the data substrate underneath it. That substrate determines whether the personal AI is a product the user owns or a product the user is.

I built Likewise because I want the substrate version of that question to have an answer that isn’t “whichever vendor shipped first.” A protocol is the layer that makes the answer adoptable by parties I will never meet. Publishing the spec before the reference implementation is, for that reason, the only sequence that made sense to me. If the shape is wrong, I would rather find out before there is an installed base depending on the wrong shape.

Please read it, especially the motivation and comparison chapters. The most useful eyes for me are anyone who has worked on Solid, the AT Protocol, Iroh, or the Willow Protocol, and anyone in personal-AI work who is thinking about the data substrate. If something you have built or argued about lands inside one of the design choices and the spec hasn’t acknowledged it, I want to know.

You can reach me at [email protected], reply on the Hacker News thread, or DM me on Twitter at @danielrmay. The repository is at github.com/danielrmay/likewise.