Research — not a product page
Nothing on this page is for sale.
Nothing described here is available to buy, and nothing on this page is a claim about what our systems do today.
An agent that has worked on a codebase for six months should know it better than one that opened it this morning. Today it does not. Every session starts from nothing, rediscovers the same structure, and pays for that rediscovery in tokens and in mistakes.
We think the missing layer is not a bigger context window or a better index. It is persistence with structure: what an agent read, what it did, what came of it — kept as part of the model rather than as a log beside it.
Our own workspace is the first laboratory, and the question is deliberately falsifiable: does an agent with six months of accumulated memory outperform a fresh one on the same tasks? We already have the instrument to answer it — the same harness we use to qualify agents — and the honest answer today is that we have not run the experiment yet.
If that question interests you more than the answer we hope for, we would like to hear from you.