Memory became the defining feature of AI products this year. The major assistants now carry context across everything, your mail, your documents, your history, and the pitch is the same everywhere, the more it remembers, the better it serves you. Researchers this spring gave the catch a name, a personalization paradox, the feature users value most is the one they can least see, audit, or constrain. Most of that conversation is about adults. I want to write about the version of it I live with, because I am building an adaptive learning platform, and adaptive learning is a version of the memory problem with unusually high stakes and unusually little attention.
Here is the bind, stated plainly. The entire pedagogical value of my platform lives in remembering what a learner struggles with. And a record of what a child struggles with is a deeply sensitive document, an authoritative, machine-maintained profile of a seven-year-old's weaknesses. The product cannot work without it and the product is dangerous with it. Both of those sentences are true at once, and this post is about living between them.
What forgetting is
Three different disciplines lay claim to forgetting, and on this platform all three apply at the same time.
For learning science, forgetting is a feature. The spaced repetition scheduler at the heart of my serving loop is literally a model of forgetting, it predicts when a concept will fade and times the next encounter just before. Forgetting is not a failure of the learner, it is the rhythm the learning works against, and the system must model it to teach well.
For privacy law, forgetting is a right. The strongest data regimes treat profiles built on behavior as something a person can demand erased, and children get the strongest protections of all. A learning record that follows a child for years, accumulating evidence of every struggle, is exactly the artifact erasure rights were written for.
For machine learning, forgetting is a threat. Wipe the memory and you wipe the adaptivity. The system meets the learner as a stranger, serves them wrongly for weeks while it relearns, and the relearning period is itself a bad educational experience. Every bit of forgetting has a pedagogical invoice attached.
A child's right to a fresh start and a child's right to be well taught are, in this architecture, partially the same budget. I have not found a way to fully escape that, only ways to spend it carefully.
Where I drew the lines
I will describe my current lines, not as the answer, just as one builder's working positions on a problem I expect to be litigated, regulated, and rethought for years.
The memory holds patterns, not transcripts. What persists about a learner is qualitative state, which concepts are mastered, which are fragile, what kinds of frustration show up and when, the signals serving decisions actually need. The blow-by-blow of every session feeds that state and then ages out of relevance. A dossier remembers everything you did. A good teacher remembers what you need. I keep trying to build the second one.
The memory serves the child, never a market. No advertising, no data sales, no third-party analytics on learner behavior, and the structural separation I built for safety means the models that generate content never see learner data at all, there is simply no pipe between those systems. I wrote about that boundary when I drew it, early in this build, and it has paid for itself in ways I did not plan.
The guardian sees through a window, not a one-way mirror. A parent gets concept-level mastery and trouble spots, the information needed to help, and not a surveillance feed. I think a child practicing alone deserves a version of privacy even from loving eyes, the freedom to struggle without an audience, because being watched changes how safely a person can fail, and safe failure is the whole product.
And the record must be killable. Erasure on this platform is real erasure, the adaptive state dies with the account, and the cost of that is honest, the system would meet a returning child as a stranger. I accept that trade in the architecture even though I suspect almost no family will ever invoke it. A right I cannot technically honor is not one I am really granting.
Where it gets harder than my answers
Schools will stress every one of those lines. The moment institutions adopt a platform like mine, the learner profile stops being a private rhythm between a child and their practice and becomes a reporting surface, progress dashboards, intervention flags, records that follow a student between classrooms. Each of those is defensible alone, teachers need signal. Together they drift toward the permanent record with telemetry, and the drift is driven by reasonable requests, which is what makes it hard to refuse. I do not yet know where I will hold the line when that pressure arrives, and I distrust my own confidence about it, which seems worth admitting in public ahead of time.
There is also a quieter question I chew on. A profile that says this child is weak at fractions shapes what the child gets served, which shapes what the child experiences, which shapes the profile. Adaptive systems are feedback loops, and a memory error can compound into a self-fulfilling assessment. The scheduler corrects against this, every fragile concept keeps getting retested, the system keeps giving the child chances to prove the profile wrong. But I notice the assessment a child cannot see and cannot contest is the part of my own product I am least comfortable with, and sitting with that discomfort seems healthier than resolving it rhetorically.
The question under the question
How much should software remember about a child? My current answer is, as little as teaching well requires, held as close as architecture allows, killable on demand, and visible to the child's adults in the form of help, never in the form of a file. I hold every clause of that loosely except the last one.
The industry's answer, watching this year unfold, appears to be as much as possible, organized later. For adults, maybe the market sorts that out. Children cannot audit a profile, cannot weigh convenience against exposure, cannot meaningfully consent, and will live the longest with whatever we normalize now. Whatever the right amount of memory is, I am fairly sure the default should not be set by what is technically easy to retain, and right now a lot of it seems to be.