2026-09-02
In 50 First Dates, Lucy wakes up every morning with no memory of the day before. Nothing accumulates in her head overnight — every sunrise starts her over. What lets her function at all is a notebook, kept by her bedside, that someone re-reads to her each morning: here is what has happened, here is who you are today, here is the story so far. The memory doesn’t live in Lucy. It lives in the notebook.
Our reasoning machine has a Lucy at its core, and it is the same character on purpose.
At the base of the system sits a large frozen language model — a network with billions of internal numbers, trained once, long before this project began, on a huge amount of ordinary text. We never train it further. Not once. Every time our machine reads a math word problem, it hands the sentence to this frozen model and reads out its internal reaction to it — and then that reaction is gone. The frozen model itself learns nothing from the encounter. The next problem, it wakes with exactly the same weights it had before, no memory of the last ten thousand problems it helped read, no accumulated wisdom about this particular project at all. It is read once per problem, the way Lucy is re-introduced once per day.
This sounds like it should cripple the system. If the part that actually understands English never learns anything specific to the task, where does the task-specific intelligence live?
It lives in a much smaller trained head sitting on top of the frozen model — a few million adjustable numbers instead of billions, all of them free to change as the system trains, none of them touching the frozen giant underneath. All of our accumulated understanding of what math word problems look like, everything the system has learned across every problem it has ever been trained on, is stored in this small head. It is the equivalent of Lucy’s notebook: the one part of the system permitted to remember anything from one problem to the next, and it does the remembering explicitly, in an object built for the purpose, not by quietly reshaping a billion frozen numbers that were never meant to hold it.
There is a second, smaller instance of the same idea living inside a single problem. Our machine doesn’t read a sentence and answer immediately — it deliberates, cycling through several passes of attention before it commits to a diagram. Across those cycles, it needs somewhere to put the things it has already figured out, so it doesn’t have to re-derive them on every pass. That somewhere is an explicit notebook: a set of named slots the trained head writes to and reads from as it works. When the head becomes confident that a particular quantity plays a particular role, it writes that down in a slot. On the next cycle, it reads the slot back rather than re-deciding from scratch. Facts that get written to the notebook persist; everything the head only half-considered evaporates when the cycle ends, the way a thought you didn’t bother to write down is gone by lunchtime.
This is a deliberate stance against a very common alternative. Many systems keep “memory” as a diffuse pattern smeared across a huge number of weights or a long running context, with no single place you could point to and say: this is the fact, this is where it lives, this is what it means. Our notebook slots are named. They can be printed out. If the system commits to the belief that a certain number represents a rate rather than a total, that belief exists in a specific slot, at a specific point in the deliberation, and you can look at it. Memory that you can read is memory you can audit — and audit, for a system whose entire second half is a wall of witnesses checking each other’s work, is not a nice-to-have. It is the whole point.
Lucy never gets to keep her memories. But she gets to keep her notebook, and it turns out that’s enough to build a life on. Our machine never gets to keep its frozen model’s memories either. It keeps a notebook instead, and it turns out that’s enough to build a mind on.