Every assistant you have ever used is brilliant for the length of a conversation and then gone. Tell it the clause that matters, the number you agreed on, the name of the person you're meeting — and a few thousand words later it is guessing, fluently, in the shape of remembering.
That isn't a bug in one product. It is what today's models are: machines that rent attention over a buffer that fills, blurs in the middle, and ends. Bigger buffers and databases bolted on the side don't fix it, because the forgetting was never fixed where it lives.
A mind you can trust with your work has to remember it. Exactly. That is the problem Kymnera exists to solve.
Human memory keeps the gist and loses the words. Frontier models do the reverse — keep the words in view and lose the middle. Kymnera does neither. As it reads, what it reads is written into memory cells inside the model — nothing compressed, nothing overwritten. Close the book and it can still give you line 4,000, exactly, because it never needed the book open.
Two parts make that a memory instead of a pile. A filing cabinet that stores everything and finds it by matching — that is what holds every result we have set. And a highlighter that marks what matters, so that however much has been read, finding the right thing stays instant and exact. Together they are the difference between a memory that fills up and one that never does. That is the memory we build.
The window slides across the document. Everything outside it is out of view; what's in the middle blurs.
The window moves on, but nothing goes dark: everything read stays as a cell. The bright ones are what the highlighter marked as mattering.
Kymnera builds models that hold the entire project in mind — every document, every decision, every number, kept word for word. Models that decide what matters without being told, answer in your words, and say "you never told me that" instead of inventing something plausible. Memory that is exact is memory you can trust with the things that matter: contracts, code, records, your own history.
That is why we started with the memory and not the chatbot — and why everything we publish is measured before it is claimed.
Sixteen thousand words later — 127 times further back than the model can see at once — it answered exactly. The same question with the memory switched off: a guess. Forty-eight times, on documents it had never seen, at spots nobody chose. Forty-eight exact. Bring your own document and see for yourself.
Kymnera-100M, September 2026. Results published, architecture withheld.
Today's models are scaled around their problems, not past them. We fix them where they start — in the architecture, from the first step — instead of paying for them at scale.
Kymnera Labs is an independent research lab founded in 2026 in Valleyview, Alberta — a solo founder building end to end with an AI research partner, on receipts: every experiment pre-registered, every failure autopsied and logged, every claim scoped to what was actually measured.
Researchers, partners, and early testers: write to us. Demo access on request.