When the new model came out, I did not switch right away. The reason was music.
Every day, I ask an LLM for music recommendations. What began as curiosity became, over a few months, the most useful recommender I had ever used. Songs I had skipped. Artists I had told it not to suggest again. Small corrections like that accumulate until a generic system becomes yours. But that accumulation was not portable. For a while, I kept the old assistant open in another tab, because only one of the two remembered that I love Sardinian choral music but do not need more of it recommended.
Eventually, I closed the other tab. After three weeks with the new model, I opened its memory file.
What I found was the most accurate description of me anyone had ever written.
Work history. The projects I keep returning to. The instruments I play. The kinds of arguments I find interesting. My family. What I am carrying this year. Where I come from, and what that means. I do not think my closest friends of twenty years could write that document. They would get the mood right and miss the facts, or get the facts right and miss the shape. The memory file had both. Three weeks. No survey, no onboarding.
This is not like the user models I have built before. A retailer's recommender may suggest another refrigerator to someone who just bought one. The math is correct. In a narrow sense, it is maximizing expected value. And still it is stupid in a way no one who knew you would be. This memory file was not stupid in that way. It read like something written by someone who knew me.
The honest first reaction is alarm. A tool I talk to casually had, in three weeks, produced an account of me coherent enough that the people who love me could not have written it in twenty years. I accept it anyway, because I get something back. When I ask about a bug, the model knows which project I mean and what I have already tried. When I ask about music, it knows what I have been listening to and which way my taste is drifting. I do not have to set the scene.
The closest analogy is not software. It is a medical record. It is personal in the same way, built up over time, painful to reconstruct from scratch, and specific enough that you want a say in who can see it and what happens to it next.
When you speak openly to a tool like this, you create a body of conversation that cannot be reproduced anywhere else. I want to be able to change my mind about which vendor I trust without starting over. That is something I have never wanted from software before.
Memory techniques will keep improving. But the layer worth preserving is the conversation history beneath them. Many frontier models can already be used without being tied to a particular vendor's memory system. If the history lives with the user, an agent can summarize it, index it, and pass only the relevant parts to whichever model is best at that moment. The user is not locked into one company's retention policy, or its decisions about whether the data gets used for training. The pieces exist. The product that combines them does not.
In coding, people are beginning to notice that the agentic harness around the model can matter more than the model itself. The same shift is coming for personal use. A system that knows the most about you, and has the best tools for using that knowledge, will be more useful than a slightly smarter model that does not know you.
Memory is where the next round of differentiation will happen. That is why this layer should live with the user.