Memory is All You Need.

The message arrived early in the morning. Subject line: "Congrats on your new role at Anthr-oogle-AI" (no relation). The big boss had decided I was the person to point the money cannon, so I sat down to write what the next useful models should actually be made of.

I did not start with a new attention variant. I started with memory.

Chat products like ChatGPT added the ability for the model to keep notes on you. On a single user that feels like a parlor trick. The assistant remembers your dog's name and your preference for terse answers. Cute. At the scale of a company, the same pattern is the first version of business intelligence that does not make you open a dashboard and squint.

Most "enterprise AI" still works like a very expensive intern on day one. No memory of the last incident. No memory of the decision you already made. No memory of the customer who churned because we promised a date we missed. Every session starts from a system prompt and a prayer. Then we blame the model when it contradicts last week's email.

A useful system writes things down.

Not a transcript dump. A memory. Something structured enough to retrieve, contradictory enough to need resolution, and editable enough that a human can say "that is wrong" without retraining a 70-billion-parameter creature. The write path matters as much as the read path. If the model cannot form a clean note, you will retrieve sludge. If it cannot forget, you will ship yesterday's bug as a fact forever.

I have been building consumer products where this is not abstract. A posture score is a memory of a body on a given day. A household pad is a memory of who already took out the trash. A coach that does not remember the serum that burned you last month is not a coach. It is a slot machine with better copy.

The companies that get this will stop asking the model to be clever in a vacuum. They will ask it to be consistent with a file it can point to. "We tried azelaic in March. Redness went down. You stopped because of the texture." That sentence is worth more than another point of MMLU.

There is a research version of this and a product version. The research version is long-term memory, continual learning, whether weights should change as you live with a model. I care about that. I also think we will get 80% of the value from a boring store, a decent retriever, and a model that is honest about what it wrote down. Continual learning can wait until the filing cabinet works.

The hard parts are not romantic:

• What is worth writing? If you write everything, you have logs, not memory.
• What happens when two notes disagree?
• Who is allowed to see a note? Memory is how products accidentally become surveillance.
• How do you evaluate that the memory helped, instead of just growing?

I would fund those questions before I funded another general chatbot that smiles and forgets.

If I really had been hired as head scientist for a day, I would have put a surprising fraction of the lab on memory: the write path, the conflict path, the forget path, and the evals that prove a returning user has a better week than a new one. Pretraining can be bought. A system that knows your organization, and can be corrected when it is wrong, cannot.

Attention was never all you needed. It was the computer. Memory is the job.