Your AI doesn't need a longer memory. It needs an ontology.

Every AI assistant you've used has the same problem. You explain your company on Monday. On Tuesday you explain it again. By Friday you've written the same paragraph about your constraints six times, and the answers still miss something you said in the second conversation.

The usual fix is more memory: longer context windows, saved chats, a growing list of "things the assistant knows about you". It helps a little. But a longer transcript is still a transcript. It records what was said, not what is true, what replaced what, or which facts belong to which part of your life.

What an ontology adds

An ontology is a plain idea with an intimidating name. It's a record of what the important things are, how they relate, and why they matter.

Take a board meeting on 14 March. As a sentence in a chat, it's just text. As part of an ontology, it's a deadline object. It belongs to a project, the Series B raise. That project is led by a person, limited by a constraint (dilution under 20%) and supported by a document, the data room index. When the date moves, one fact changes, and everything that depends on it can see the change.

That structure is what lets an AI answer the question you meant rather than the question you typed. "Are we on track?" is unanswerable from a transcript. From a connected record, it's a matter of looking at the deadline, the decisions taken, and what's still open.

Why structure beats volume

Three things get easier when context is structured rather than piled up.

Relevance. A conversation about your product launch shouldn't drag in your house renovation. When each fact belongs to a World, the context a model receives can be scoped to the work in front of it.

Correction. When something stops being true, you need to be able to set it aside without deleting history. A connected record can keep the old valuation for reference while keeping it out of today's answers.

Inspection. If an answer feels oddly generic or built on the wrong assumption, you should be able to see exactly what the AI was working from. You can't audit a vibe. You can audit a list of objects and their links.

How Manwe does it

In Manwe, each company, project or area of life gets its own World, and larger ones can have Moons for their parts. A World holds the people, goals, facts and constraints around that work.

Memory has three places a fact can be:

• Memory: retained context Manwe can use in the right scope.
• Review: proposed context waiting for your decision. It never enters an answer until you approve it.
• Set aside: kept for reference, excluded from answers, and restorable.

When you mention a deadline or a decision, Manwe can remember it straight away and show a Remembered note with Undo under the answer. Material from documents, reviews or other AI tools goes through Review first. And the memory map shows what's retained and how it connects, so the structure is something you can see, not just something you're told exists.

This is early work, and we say so in the app. Memory and learning are improving quickly, and the ontology is the part we're most careful with, because it's the part that makes everything else either useful or confidently wrong.

The point

Models will keep getting better at reasoning. What they'll never have, on their own, is your world: the people, the commitments, the constraints and the history behind every question you ask. That has to live somewhere you control, in a shape the model can use.

That's the job of the harness. Not a longer memory. A better-organised one.