Memory

A school keeps two things about a child with any reliability: a sequence of marks, and a record of promotion. Both are administratively necessary. Both are close to useless to the person who has to teach the child next.

Consider what a mark carries. A term score of 43 records that a child lost 57. It does not record which 57. It does not say whether the loss was one misunderstanding repeated across six questions or six unrelated gaps, whether the child ran out of time on work they could have done, or whether they answered a different question from the one asked. Two children with identical marks can need opposite things from the next teacher. By the time anyone might want to look properly, the script has been returned, or filed, or thrown away.

Now consider what the teacher knew. Over a year a teacher builds a working model of a child. That this one inverts the sign when moving a term across the equals sign, but only when the term is already negative. That this one cannot begin a word problem until somebody reads it aloud, and then solves it correctly. That this one has decided they are bad at geometry and will not attempt a construction at the board, though they will at their desk. That the explanation in terms of money finally worked for this child, after two years of explanations in terms of the number line did not. That this one’s silence means confusion and that one’s silence means boredom. That asking this particular child a question in front of the class ends the lesson for them.

None of it is written down. It is not written down because there is nowhere to write it that anybody will read; because writing it would take longer than the teacher has, across five sections; and because much of it is held as instinct rather than as sentences the teacher could produce on demand.

Then the year ends. Or the sections are rebalanced in July, or the teacher takes leave, or is promoted out of the classroom, or moves to another school, or the family moves city and the child arrives in March with a transfer certificate and a mark sheet. Whichever it is, the model leaves with the teacher, and the child arrives at the next desk as a stranger with a number attached.

The cost lands in three places. The next teacher spends a term rediscovering what was already known, and with a full section will not finish the rediscovery. The approaches that already failed are tried again, because nothing recorded that they failed, so the child sits through the number-line explanation for the third year running. And the child pays the largest share, because they must present themselves again each year to somebody who has no account of what they understand, and in the absence of such an account a slow start reads as a limit.

There is a quieter loss. What a teacher retains is biased towards what is memorable: the very strong, the disruptive, and the visibly failing. The child who is two years behind in one specific thing and otherwise unremarkable is held by nobody. That difficulty is invisible in the marks, invisible in the staff room, and survives every handover intact, which is why it is still there in Class 11 wearing a different name.

This is not, in origin, a technology problem. It is a problem of what institutions choose to retain, and schools retain what they are required to report. But the binding constraint is the absence of any durable, usable account of a learner that outlives the adult who formed it. That absence is what this line of work is about. It is not something we have built a remedy for.

The obvious wrong answer is a record of activity. Knowing how many questions a student attempted, how many hours they spent in an application and how many videos they watched is not knowledge about that student. It is knowledge about the application. A log of behaviour feels like memory because it is voluminous and time-stamped, and it answers none of the questions a teacher actually has.

The useful thing is an account of understanding: what is known, what is shaky, what is blocking what, and what has already been tried and did not work. The test of such an account is not how much it holds. It is whether a teacher who has never met this child can ask it something on a Monday morning and get an answer they can act on before the bell.

So: which of the three things I could teach this week is actually blocked, for this child, by something they should have learnt two years ago? Where is this child right often enough to pass and wrong in a way that will surface again in a public examination? What has already been tried with this child, by whom, and what happened? Is this difficulty in the mathematics or in the reading of the question? What does this child do when they are stuck, and what does that look like from the front of the room? Which explanation has worked for this child before, as distinct from which explanation works for children in general? What did this child appear to understand last term that they no longer appear to? And the question a teacher would actually ask if the previous teacher were standing in the corridor: what is the one thing you would tell me about this child in ninety seconds?

Two further properties are not optional. Such an account must be able to say that it does not know, and mean it, distinguishing what a child has not demonstrated from what a child cannot do. And a teacher must be able to see what any claim rests on, and when it was formed, so that they can decide to discount it. An account that presents everything at the same confidence is worse than no account, because it invites a teacher to act on the weakest thing in it.

Nothing here is a statement about how such a memory would be held. That is deliberate. The hard part of this problem is agreeing on what it must be able to answer, and for whom, and under what permission. The rest follows from that and is much easier to argue about.

Almost all attention in this area goes to the learner. The neglected half is the teacher, who accumulates over a career an enormous body of practical knowledge and leaves almost all of it behind on the way out.

What a good teacher knows is specific: which explanation lands against which misconception, and in what order; which question to ask first to find out where a class actually is; how to sequence a chapter for a weak section without abandoning the strong students in it; which past examination item exposes a particular gap cleanly; what to say to a parent who has arrived angry; when to leave a child alone. This is a genuine professional asset, built slowly, and the handover on retirement or resignation is a syllabus tracker and a mark sheet. The school keeps the topics covered and loses the craft that covered them.

We hold that a teacher’s memory is as much a research object as a learner’s, and that treating it as an afterthought is a mistake about where the value sits. It is also harder, for three reasons.

First, the knowledge is tacit. Teachers express it as judgement, not as rules, and asking a teacher to write down what they do reliably produces something more generic and less useful than what they actually do.

Second, it is entangled with particular children. An explanation that worked is an explanation that worked for that child, in that room, in that week. Separating the durable craft from the local accident is the whole difficulty, and we do not have a principled way to do it.

Third, and most seriously: a record of how a teacher teaches is a record that can be used to appraise them. Anything rich enough to be worth keeping is rich enough to support ranking, performance management, contract non-renewal, and comparison between colleagues who teach different children. We treat that as a reason the work may not deserve to exist, not as a risk to be handled in a later phase. And there is a failure that defeats the exercise from the inside: a teacher who believes their practice is being recorded for appraisal will curate what is recorded, and a curated record of practice is worse than no record, because it is confidently wrong. Whatever control a learner has over what is held about them, a teacher needs at least as much, and probably more, because on the other side of a teacher’s record sits their employment.

State the position without hedging. A memory of a learner that exists only inside one company’s system is a reason not to leave that company. That is its commercial function, whether or not anybody intended it. The longer such a memory runs, the more it costs to leave, and the cost is paid by a child in the form of a history that does not follow them, as a consequence of a procurement decision made by adults for reasons that had nothing to do with them. Building a switching cost on top of a child’s education is indefensible, and we would rather say so on a page than defend it later.

The guarantees that follow are the position, not a policy annexe. The memory belongs to the person it is about, not to the school that generated it or the company that holds it. Access is granted rather than assumed, and granted for a stated purpose rather than in general. Access is revocable, and revocation has to be real rather than a setting that merely stops the display. The memory can be exported in full, in a form that is useful somewhere else and not only here, including to a competitor. And the person can see what is held about them, in language they can read without a technical intermediary, including the parts that are unflattering.

A related initiative at altme.org pursues the same position for people generally rather than for learners: a person’s own memory, held privately, portable, and granted to an agent with the person’s consent rather than accumulated silently as a by-product of using something. It publishes a draft open protocol, version 0.1, as a request for comments, for portable person-owned memory, together with a reference implementation. It is a draft. It is not ratified by anyone. There is no adoption to report, and we will not present a document nobody has implemented as an ecosystem. How any of it works is not the subject of this page.

Portability is a commercial disadvantage. That is what makes it a commitment rather than a feature, and it is also what makes it checkable: if we ever quietly abandon it, the visible symptom will be that the export stops being useful. Anyone who cares whether we mean this should test that, periodically, rather than believe this paragraph.

The strongest argument against this work is not that it will fail. It is that it will work as described, and that a durable account of a child is a durable account of a child.

A judgement formed in Class 6 arrives in Class 11 ahead of the person it is about. The child who was slow at eleven and is not slow at sixteen walks into a room where the account has already been read. The teacher who reads it is not being careless. Reading it is the point, and it is exactly what this work would ask them to do. But what is read before a child is met becomes what the reader watches for, and a child who is treated as the account describes has few ways to produce evidence against it. Nobody in that room does anything wrong, and the record still closes over the child.

Human memory is unreliable, and one of the things its unreliability does is let a child stop being who they were. A teacher who half-remembers can be surprised. Reputations fade in a staff room because staff rooms forget. This work proposes to remove that forgetting on purpose and to describe the removal as the benefit. We do not know how to give the mercy back without discarding the part that is useful, and we have no candidate for how it might be given back.

The available answers are weaker than they sound. A settings page moves the burden onto a fifteen-year-old to know what is held about them and what it would cost them to remove it. Deletion on request is a right exercised by the confident and the well-advised, who are not the same population as the harmed. And a record built to help a teacher teach is a record available to whoever else can open it: for setting and streaming, for a scholarship decision, for an admissions file, for a reference written by somebody who never taught the child. Records about people do not stay inside the purpose they were built for. Assuming this one would is the least defensible assumption in the whole idea.

The same argument runs on the teacher’s side, where what sits next to the record is their employment.

We think a serious person could read this section and conclude that this should not be built at all. On the argument as it currently stands, that is a reasonable conclusion, and it stays reasonable until somebody has an answer better than an assurance of good intentions. We do not have one. Nothing here has been tested, so this is reasoning and not evidence; the risk does not need evidence to be taken seriously, and a promise to be careful is not an answer to it.

Arguing for something with no results is only honest if you say in advance what would count against it. Here is what would tell us that a persistent learner memory is not worth having, roughly in order of how likely we think each is. The entrenchment argument in the previous section is not on this list, because it is not a way the work fails; it is a way the work succeeds and does damage.

Teachers do not use it. This is the most probable failure and the easiest to conceal. A teacher with a full section, a syllabus to finish and forty minutes has a very high bar for consulting anything, and an account that is not consulted before the lesson does not exist. If teachers are given a durable account of a child and it changes nothing about what they do on Monday, then the artefact is for us and not for them, and a well-run demonstration will hide that for a long time.

It goes out of date faster than it accumulates. Children change quickly, and an account whose contents stop being true faster than new observations arrive is not neutral. It is a liability: confidently wrong about a child who has moved on, and wrong with the authority of a written record.

It distributes the benefit backwards. If a learner memory helps most in the classroom that already has an attentive teacher with time, and helps least in the section of sixty with a substitute in charge, then it widens the gap it was meant to close. The aggregate can look good while this happens, which is why an aggregate result would not settle the question.

And the cheap-alternative test. If the same account could be obtained by giving the next teacher fifteen minutes with the previous one, then none of this earns its cost, and the intervention worth funding is fifteen minutes.

We would rather find these things than not find them, and we currently have no evidence in either direction, because nothing has been measured.

These are unsolved, and they are not unsolved in the sense of being unimplemented. We do not know the answers, and in several cases we do not know what a good answer would look like. Anybody who tells you these are settled is selling something.