8 October 2026 · how it works
How AI memory works: what you own, what the model owns, and where your two years of training actually live
By the end of 2028, enterprise AI contracts will carry a clause about who owns the stored context when a deployment ends. The weights fight is already happening. This one isn't yet. Bet it.
Your AI forgot you. You switched apps and it started again from zero. Not a bug. The architecture.
There are three things called AI memory. They don't live in the same place. They're not owned by the same party. The one you think you're building is the one you have the least claim to.
What the model knows: weights
When a model is trained, it adjusts billions of numbers.1 Those numbers are the weights. They freeze when training ends. You can't change them without retraining the whole thing.
This is the model's long-term memory. It's shared across every user on earth. It belongs to the company that trained it. A full-precision copy of Llama 3.1's 70 billion parameters weighs 140 GB.2 That 140 GB contains everything it learnt before you typed a word. You have no copy. No claim. Changing a single fact inside it requires retraining, or risky patching that can break other things.3
What it holds during your chat: the context window
Every word you send, every reply it gives, goes into the context window. Think of it as a scratchpad, wiped when the session ends.1 Llama 3.1 holds 128,000 tokens4, roughly a long novel's worth. Gone when you close the tab.
Here's the number none of the sources put side by side. The key-value cache for a Llama 3.1 70B session at full 128,000-token context runs to about 1.3 GB.2 The weights are 140 GB. Divide one by the other: 108 to 1. What you said in this chat is roughly one per cent of what the model holds, measured in storage. When the session ends, that one per cent is gone. The 140 GB serves the next person.
What platforms call memory: notes in a database
When ChatGPT or Claude says it remembers your job title and your tone, it has written a note to itself in a database. Not in the model. Systems like these hold user profiles and key-value stores outside the weights entirely.5 Next time you open a chat, those notes get loaded in at the start.
This is where it gets commercially interesting.
In March 2026, ChatGPT, Claude, and Gemini all shipped memory portability features within weeks of each other.6 You can now export a text list of what the platform wrote down about you. Claude gives you your full profile: tone, personal details, projects, corrections.7 You can paste it into a rival.
But the export captures stated preferences, not the patterns that built up across thousands of actual chats.8 The import tools make a one-time copy. After that, each assistant's notes drift apart.6 Portability hands you a photograph of the relationship. Not the relationship.
Who this costs
Not the person who knows about the export button and uses it. Her.
The person on month eighteen with a well-tuned assistant. The one who doesn't know the export exists. Who assumes the relationship is durable. Who finds out, when the employer switches vendor or the platform changes pricing, that memory was a tenancy, not a deed.
The note about her job title lived on the platform's servers. The platform decided the format. The 140 GB stayed exactly where it was. She got a text file.
The weights question, who owns the model itself, gets all the legal attention right now. The stored-context question, who owns what a model built up about this organisation over two years, is the one nobody has litigated. It's coming, and it will cost more when it arrives.
The refutation
The load-bearing claim: what you export is not the model's memory, it's a note in a database. The search that tries to break it: does any deployed product write your data back into the weights themselves?
The answer right now is no. A single set of weights has no way to tell one user from another at the parameter level.3 Per-user weight updates exist as a research idea, not a shipped product at consumer scale. The claim holds for every product you can name today. If that changes, the text-file export misses the main event, and the ownership fight arrives faster.
They kept the 140 GB. You had the 1.3 GB. They sent you the notes.