The tutor's memory of the student is a list the student can read

Each student has a tutor agent in their own container. Over a semester it remembers things: this student skips tests, that one confuses recursion with iteration. Memory is keyed by project: one project per student. Tokens go through the Gateway from the department's AI Balance; the memory stays on the box.

The student can see it. GET /api/v1/agent/memory/items?project=student-17 lists every record. PATCH corrects one; DELETE removes it. PUT /api/v1/agent/memory/enabled turns capture off for exam week. The consolidation pass that merges records is human-only: the agent cannot approve it for itself, so what it remembers never rewrites what it remembered.

A hosted tutoring product keeps its model of the student on the vendor's side. Here the model of the student is data in the student's container, on the department's server, and the API to inspect it is the one the agent uses.

What the tutor knows about you is a GET, and forgetting is a DELETE.