An AI Tutor That Remembers Last Year: The Pieces Exist, the Trial Does Not

The problem

Starting point. The earlier paper in this series found that student-facing generative tutors had been trialled in Nigeria, Turkey, Ghana and the United States. Structured, supervised programmes showed gains of roughly 0.2 to 0.4 standard deviations, while the largest trial of an off-the-shelf tutor found gains no bigger than the same platform delivers without AI, because most students rarely used the tutor¹. Almost all measured deployments were research pilots rather than system-wide services¹. This paper asks something different: what is lost when a student changes teacher or school, and whether a tutor could carry the student's learning history across that point.

For most students a new teacher each year is the normal case. In Tennessee's statewide records, only 6% of students in a given year had a maths or English teacher who had taught them before, and fewer than 2% of teachers worked in intentionally looping classrooms, where a teacher stays with a class for a second year (administrative data, grades 3 to 11, 2007 to 2017, working paper, June 2022)².

The formal handover is thin. In England, schools maintained by a local authority must send a Common Transfer File (CTF) to a pupil's new school within 15 days of the pupil leaving³. Its modules cover identifiers, free-school-meal and special-educational-needs history, attendance, results of completed key-stage assessments and the pupil's school history³. The guide presents the file as a way for the new school to plan around prior attainment and set targets, and notes that missing performance history can harm target setting³. None of the listed modules records how a pupil works through problems, which misconceptions recur or what kind of help worked. That is a reading of the module list in the 2020 guide, not a statement made in it³.

The cost of a lost history is real but small on average, and uneven. In a Tennessee study of 1.3 million students (June 2022, observational, student fixed effects), having a teacher for a second year was associated with test scores 0.02 standard deviations higher, with fewer suspensions and, in high school, fewer truancies². The authors cannot separate the teacher adapting to the student from the student adapting to the teacher². A study of six California districts (grades 4 to 8, 2016-17 to 2019-20, observational, October 2025, abstract) found near-zero effects of scheduled and summer school changes, but mid-year moves lowered English scores by 0.073 and maths scores by 0.083 standard deviations, with the largest losses for students with disabilities, English learners, low-income students and homeless students⁴.

What already works

The building blocks exist separately.

Tutors that help in structured settings. In a randomised trial among volunteers in Nigeria, students offered a six-week tutoring programme scored 0.31 standard deviations higher than a control group that had no extra sessions, so the result combines the tutor, teacher facilitation and about 18 extra hours of study (May 2025 working paper)⁵. The two-year cluster randomised trial in Tennessee (August 2026) found gains resembling those from the same platform's practice without AI assistance, and the median student messaged the tutor on a third of practice days⁶.

A record the student can read. Khan Academy's help centre describes a History tab where a student can read past tutor chats, preview a short summary of each and resume a chat (operator documentation, updated July 2026)⁷. The page does not say whether the tutor draws on earlier chats when it responds, or whether anything survives a change of school year⁷.

Memory in a research prototype. A framework in an AAAI 2026 paper keeps a structured profile of each student's proficiency, a memory of past interactions and an estimate of forgetting over time. It reported an average improvement of 11.4% in normalised learning gain over the strongest baseline across four public mathematics datasets (simulation study, date not shown)⁸. The tutoring sessions were played out with simulated students built from real learning logs, personalisation was scored by an automated judge, and the authors list deployment with human learners as future work⁸.

A record that already crosses the boundary. The CTF shows that moving a pupil's history between schools is routine and legally required in England, through a secure transfer system run by the Department for Education³. The same guide lists cases where passing the record on may be inappropriate, including a family escaping a violent partner, a family in witness protection and an adopted pupil³.

No trial was found in which a tutor's record of a student follows that student into the next school year, and none in which a receiving teacher or tutor uses such a record. Tutors with measured gains, memory in prototypes and a legal handover of pupil records exist separately. Nobody has shown them working together.

The possibility

One way this could work starts at the end of a school year. A tutor that had worked with a student over the year could draft a short summary in plain language: what the student had practised, which misconceptions kept returning, what kinds of hint helped and what the student could do without help. The summary would contain no grade and no prediction. The student's teacher would review it, correct it and decide what goes forward; the student and parent would be able to read it and ask for it to be deleted. The next year's tutor would begin from the summary instead of a blank diagnostic.

A variant would serve the group where the evidence of cost is largest. If a student moved school in mid-year, the summary could travel alongside the transfer of records, so that a new school that knows nothing of the student receives more than results and attendance. It would be a second document next to the CTF, not a change to it.

A narrower first step would keep memory inside one school year, across terms, where custody and retention risks are smaller and an evaluation is easier to run. If that worked, the year boundary could become the second test.

A summary that follows a student into a new classroom could help the new teacher or prejudge the student, and no study has tested which.

What would have to be true

1. Students would have to use the tutor. The median student in the Tennessee trial messaged the tutor on a third of practice days⁶. A profile built from sparse use would describe the student's habits more than the student's understanding.
2. A carried-over profile would have to beat a fresh start, and simpler handovers. No trial tests this, and the memory evidence comes from simulated students⁸. A teacher-written note, or keeping a class with the same teacher, needs no new technology. A teacher who already knew the student was associated with an average gain of 0.02 standard deviations². If a tutor profile produced effects of that order, they would be hard to detect in a trial and would have to justify their cost; the Nigerian six-week pilot cost about USD 48 per pupil (May 2025 working paper)⁵, and no cost figure for keeping a profile across years was found.
3. The record would have to describe, not label. Teacher expectations have been linked to educational attainment, with an estimated elasticity of college completion with respect to teachers' expectations of about 0.12 (US data in which two teachers rated each student; working paper, November 2018; observational)⁹. Teachers were on average overly optimistic about whether students would finish a four-year degree, and the over-optimism of white teachers was larger for white students than for Black students⁹. A summary that lowered a new teacher's expectations could do harm, and one that raised them could help, so its content would have to be confined to observed work, open to correction and free of predictions.
4. Custody and retention would have to be settled. In the PowerSchool breach (December 2024), data on 270,812 students in one Canadian province was accessed, about 75% of them no longer in the school system, and the province says retention rules in its Schools Act prevent it from deleting the records (provincial statement, updated June 2026)¹⁰. PowerSchool paid a ransom on assurances that the data would be destroyed, but the data was not deleted, and school districts in Canada later received extortion demands using it; the province found no evidence of such attempts in its own schools¹⁰. In the United States, the amended children's privacy rule (COPPA, January 2025) bars covered operators from retaining children's personal information indefinitely¹¹. A tutor profile would need a named holder, a defined lifetime and a deletion route that survive a change of vendor.
5. The boundary would have to be able to stop. The cases in which the CTF guide says a record should not follow a pupil would apply equally to a tutor profile, and the profile would need the same exits³.
6. The legal category would have to be settled. Annex III of the EU AI Act lists as high-risk the AI systems intended to evaluate learning outcomes, including when those outcomes are used to steer the learning process¹², and the rules for those systems apply from 2 December 2027 after the Digital Omnibus on AI¹³. A tutor that keeps a running estimate of what each student has mastered and uses it to choose what comes next sits close to that wording. Whether a given product is covered depends on its intended purpose, and that classification was not verified.

What could stop it

What is documented. The benefit of continuity is small on average. The estimated gain from a second year with the same teacher is 0.02 standard deviations², and scheduled and summer school changes show near-zero effects⁴. The largest losses come from mid-year moves, concentrated among students with disabilities, English learners, low-income students and homeless students⁴. The earlier paper found that in the US trials the students who used the tutor least tended to be lower-performing¹.

Records have been breached, and central stores have provoked backlash. The PowerSchool incident above is one case¹⁰. Another is inBloom, a US venture backed by USD 100 million from two foundations. It planned to hold more than 400 data fields per student, including family relationships and reasons for enrolment changes, and it shut down in April 2014 after parent protests and a New York law restricting outside companies from storing student data; parents in Louisiana found their children's Social Security numbers in the database (news report, April 2014)¹⁴. inBloom stored administrative data and was not a tutor, so the comparison is imperfect. The trust question it met, which data goes where and who controls it, is the one a cross-year tutor profile would meet.

The official plans do not mention the idea. The Department for Education announcement of January 2026 describes AI tutoring tools for up to 450,000 disadvantaged pupils in Years 9 to 11, with trials in secondary schools later in 2026 and tools available to schools by the end of 2027, designed to complement face-to-face teaching and to build on children's learning in class (government plans, not results)¹⁵. As opened, the announcement does not mention carrying a pupil's record across years¹⁵.

What is reasoning. Incentives could work against it. A profile held by a vendor ties a school to that vendor's product, and a profile held by the school brings a security burden the school may not be able to carry. Those costs would fall on schools and families, while the benefit, continuity, is small on average. The students with most to gain from a carried-over record may also be those least likely to have built one, if low use persists.

Simpler solutions could do much of the job. A teacher's handover note, a repeat-teacher or looping assignment, or a fuller transfer file could supply continuity without a machine-held profile. If the value of continuity comes mainly from a teacher knowing a student, a record may be a weak substitute for knowing a student, and a cross-year tutor could deliver less than it first appears to.

Signs to watch

• Trials of the Department for Education's AI tutoring tools in secondary schools, announced in January 2026 for later that year, with tools for schools by the end of 2027: whether any trial follows pupils across terms or years¹⁵.
• 2 December 2027: rules for high-risk AI systems in Annex III, which includes systems that evaluate learning outcomes to steer learning, apply¹²¹³.
• An independent trial of the redesigned Khanmigo interface, which Khan Academy says is built to push students to engage¹.
• Estonia's AI Leap expansion in 2026/27, with University of Tartu researchers measuring long-term impact; no results date is published¹.
• A real-student evaluation of a memory-based tutor of the kind described in⁸, which its authors list as future work.
• Any published trial in which the same students keep or lose a tutor's record of their work across a school-year boundary; none was found.