Written once. Frozen forever. Recalled exactly.
Every frontier model rents attention over a buffer. The buffer fills, decays, and ends — and the model forgets. Kymnera puts memory in the weights: a native memory organ, trained from the first step, that stores what it reads and gives it back exactly.
Try the demo — bring your own documentContact108M parameters, byte-level, trained from scratch. Results only; architecture withheld. Every run pre-registered, every result on an append-only ledger.
Today's models are scaled around their problems, not past them. Forgetting, hallucination, and cost get patched from the outside — bigger buffers, retrieval scaffolding, more GPUs — and remain problems.
Kymnera fixes them where they start: in the architecture, trained in from the first step. Memory in the weights. Honesty as a trained reflex. Efficiency as a design constraint, not a bill. The goal is a frontier-class model family whose advantage is architecture, not compute budget.
The substrate. Reads a document 768 bytes at a time and lets each piece go; stores labeled records into memory cells and recalls them exactly at any distance. It remembers what it is told to remember.
The instincts. Store what matters without being told. Recall from questions in your own words. Refuse to guess when nothing was stored. Trained beside a plain transformer of the same size for the controlled comparison.
Comprehension over memory. Enough model to use what it remembers: answer questions about a document it read once; hold a long working session without forgetting or inventing.
We say which rung a result belongs to, every time.