Why a human approves everything before it enters memory

Ontonym uses language models to do the tedious part of remembering: you paste meeting notes or an email thread, and the model pulls out the facts — people, companies, dates, decisions, and how they connect. It’s good at this. Most days it’s very good.

But “very good” has a sharp edge when the output is shared.

A wrong note in your own head hurts only you. A wrong fact in a shared memory spreads. Your teammates read it and believe it. Every agent connected to the memory repeats it, confidently, in every answer it touches. The whole point of shared memory — one version of the truth — becomes the problem when that one version is wrong.

So there’s a gate

Nothing the model extracts goes straight into memory. It lands in a review queue first. A person looks at each proposed fact and approves it, fixes it, or throws it out. Only then does it become part of what the team — and the team’s agents — see.

Approving is fast, because the queue shows exactly what will change: the new object, the changed property, the link about to be created. Most items are one glance and one click. That glance is the entire difference between “the AI said so” and “we know this.”

Where a fact came from stays visible

Every fact keeps its trail — what it was extracted from, and that a person let it in. When something looks off six months later, you don’t have to argue about it. You look at where it came from and decide.

The pattern is worth naming, because it applies to more than us: let the model do the work, let a person own the truth. We’d rather your memory grow a little slower and stay something your team actually trusts.