An AI tutor that actually remembers how this student learns.
Personalized learning only works if the AI carries forward what it learned about each student. MemexAI stores learner profiles as human-readable files — growing with every session, correctable by teachers and parents, and never locked inside an opaque model.
Personalization that resets every session isn't personalization.
The promise of an AI tutor is that it learns how this student works: which explanation style clicks for them, where their conceptual gaps are, what they've already mastered, what they gave up on last week. That profile is the whole product. Without it, you have a smart chatbot that happens to answer math questions.
Most AI tutors accumulate this context but can't surface it. Ask the system what it knows about a specific student and you get silence or a summary that can't be verified. Ask it to correct a wrong belief — "she's actually strong at fractions, she just struggles with word problems" — and there's no clear path to make that correction stick.
For parents and teachers, this opacity is a dealbreaker. They're not just consumers of the AI's output — they're co-educators with context the AI doesn't have. A system that accepts their input but hides its model of the student loses their trust and their engagement.
The correction problem matters more in education than almost anywhere else. A wrong belief about a student — "struggles with algebra" when the real issue is notation, not algebra — shapes every subsequent session. If that belief can't be corrected by the teacher who knows the student, the AI compounds the error over time.
Scale makes this worse. A tutoring platform with ten thousand students can't rely on incidental correction through conversation. When a teacher needs to update a student's profile — they changed learning pace, they were misassessed on a skill, they have a new accommodation — there needs to be a surface where that correction happens and sticks.
A learner profile teachers can read. Students can question. Parents can trust.
What the tutor writes
As the AI tutor works through sessions, it builds a structured learner profile in memory files: skills mastered and in-progress, preferred explanation styles, recurring difficulty patterns, pacing notes. Each session adds to the profile — it never resets.
// After a math session
await memory_remember({
content: `## Learner profile — Maya
Strong: fractions, basic algebra
Working on: word problems (parses correctly, struggles with setup)
Responds well to: visual diagrams, real-world examples
Avoids: timed exercises (anxiety)
Last session: covered ratio problems — needs another pass on
multi-step ratios before moving to proportions`,
path: "user/learner-profile.md"
})What teachers and parents can do
The admin console shows every student's memory files in plain language. A teacher who knows a student's situation better than the AI can open the profile, correct a wrong assessment, and add context the AI couldn't have learned from sessions alone: "she's been tested for dyslexia, prefers audio explanations."
Revision history shows every change to a learner profile — when a skill was marked mastered, when a difficulty was first noted, when a teacher corrected something. For parents asking "how does the system think my child is progressing?", the answer is an actual file they can read.
Shared files under shared/ hold curriculum guidelines, grade-level expectations, and pedagogical policies. Update the curriculum once and every tutor session in your deployment reflects the change — no redeployment, no per-student prompt updates.