Everything a machine learns dies with it.

Physical AI is no longer a slide in a keynote. Roughly 16,000 humanoid robots were deployed in 2025. Goldman Sachs projects a $38B humanoid market by 2035. Bank of America expects unit costs to fall below $17,000 by 2030. Agility already rents its Digit robot at $10 to $12 per hour, against $30 for the human it works next to.

The direction is set: machines that see, move, and manipulate the physical world are coming into warehouses, construction sites, kitchens, and fields. The hardware is commoditizing. The models are improving every quarter. Billions of other people’s capital are making sure of it.

The interesting question is no longer whether machines will work among us. It is what happens to everything those machines learn.

A protocol stack is forming underneath Physical AI, the way one once formed underneath the internet. Transport exists: ROS 2 and MQTT move messages between machines. Training exists: VLA models and formats like LeRobot turn demonstrations into skills. Spatial context exists: OpenUSD describes the world machines operate in. Safety is forming: certified layers like Halos.

Four layers. And then it stops. There is no layer for deployed knowledge — for what a machine learns after it ships. Where does it live? Who owns it? What happens when the machine is replaced, or the vendor changes, or the fleet doubles?

Today the answer is: it dies. A robot that spent three years learning a building is decommissioned, and its successor starts from zero. Everything the stack can transport, train, and situate — it cannot remember. That is the missing fifth layer.

If the fifth layer gets built inside a robot manufacturer, the result is predictable: your machines learn your site, and the manufacturer owns what they learned. Switching vendors means abandoning years of accumulated knowledge. That is not a memory layer. That is lock-in wearing a memory layer’s clothes.

We hold the opposite position. What a machine learns on your site, about your processes, from your people, is yours. It should live where you control it — sovereign cloud or on-premise, keys in your hands — and it should transfer across vendors, across machine generations, across time.

Regulation is starting to point the same way. The EU Machinery Regulation, in force from January 2027, requires documentation of machine behavior. ISO/WD 26264-1 drafts a standard for humanoid datasets — but it covers dataset architecture only, not runtime knowledge or ownership. The standards say knowledge must be structured. Nobody has said it must be owned. We do.

Kin is the fifth layer, built as a product: one persistent knowledge entity per person, per site, per fleet. It holds what has been learned — the routines, the exceptions, the corrections, the context — independent of any single machine.

Machines connect to a Kin and inherit what it knows. A new robot on a site your Kin has known for years starts with those years, not with an empty state. What the machine learns while working flows back into the Kin, so the entity grows with every machine that touches it.

Kin builds on the standards that exist rather than replacing them: ROS 2 and MQTT for transport, VLA and LeRobot-compatible skill formats, OpenUSD for spatial context. One AI. Every machine.

A fifth layer with nothing in it is an empty shelf. The most valuable knowledge on any site today is not in machines at all — it is in people, and it is retiring. The caretaker with 25 years of one building in his hands. The best cleaner. The mason who knows which wall will give you trouble. None of it is in a manual.

Lore is the instrument that captures it. A skilled worker records while working — video and voice, with the phone in their pocket or a bodycam they already wear. Lore structures it into searchable task knowledge inside the company’s own Kin. New hires learn from it today.

We are precise about the phasing. A recording is not yet a robot-executable skill; robot-executability grows as the training standards mature. But the knowledge, once captured and structured, is ready to make that transition — and in the meantime it does the older, plainer job of not letting expertise walk out the door. Lore is the bridge between the human era of work and the machine era: capture what your best people know today, and when the robots come, your business starts at day one. Your competitor starts at zero.

Follow the idea far enough and knowledge becomes an asset class. A business that sells does not just transfer buildings and contracts — it transfers a Kin that holds decades of operational knowledge, and the buyer can verify exactly what it contains. A family firm hands its Kin to the next generation the way it once handed down the workshop. Succession stops being the moment knowledge is lost.

The word lore has always meant this: practical knowledge handed down, from the people who had it to the people who need it. Kin is the family. Lore is what the family passes on. We are building the infrastructure that lets that handover survive the arrival of machines — and outlast any single one of them.

The internet shipped without an identity layer, and the world has spent thirty years paying for the retrofit. Physical AI is shipping right now without a knowledge layer. This time the gap is visible before the wave peaks. benned exists to build that layer — before it gets built as someone’s lock-in.