We started sleds as a memory layer, and for a while that was the product. Then, over the winter, the frontier labs shipped memory of their own, and it kept getting better each month. Claude remembered the project. Agents carried context across sessions. The thing we had built by hand was becoming a feature of the models themselves.
This is the kind of moment a company either argues with or listens to. We listened.
What memory did not fix
When the model remembered everything, the same work still got stuck in the same places. A request still sat in a tracker without a done line. Nobody still knew what it would cost until it was done. The support desk still re-explained the bug to product, who re-explained it to engineering. The agent could remember the codebase; it could not get the customer's sentence to the person about to build.
The context was not lost anymore. It was still not moving. The gap was never memory. It was the seam between the people who want things and the people who make them, and the fact that nothing ran across it except meetings.
The two things underneath
Looking at what our own team actually did all day, two things stood out.
Orchestration. Agents could do the work, but someone still had to decide what ran, where, with what context, and whether it was done. That someone was a person with a terminal and a list. The list was the product, and it was in their head.
The connection to the business. The biggest gains we had seen were not from a faster agent. They were from the day a product manager could file a request that arrived understood, and the day the customer's words were on the row the engineer read. That was not a memory feature. It was intake, a size, a route, and a receipt.
So we turned. The memory layer became the pack, the context a request carries with it, and the company became the rest: where the business meets engineering.