Grounding AI in Knowledge: RAG and Memory for Reliable Tutors

The best-known flaw of LLMs in education is hallucination: producing a confident but wrong answer. In a classroom, a confidently stated error does damage. Two complementary mechanisms aim to address it: retrieval-augmented generation (RAG), which grounds answers in a reference corpus, and memory, which lets the system remember the learner over time.

RAG: grounding answers in the right content

The RAG principle is simple: before answering, the system retrieves the relevant passages (textbook, curriculum, the teacher's course) and bases its answer on them. A systematic review of 51 studies catalogues educational uses — interactive tutoring, content generation, assessment — and open challenges: reducing hallucinations, ensuring retrieval completeness, controlling compute cost and multimodality ^[1].

Results are encouraging but nuanced. A RAG tutor deployed across four courses reduced hallucinations for well-covered topics, but 1.5% of answers were still incorrect and 16.5% fell outside the provided context ^[2]. Combining multimodal knowledge graphs with hybrid retrieval reached roughly an 82% fully-correct answer rate while lowering construction cost ^[3].

Memory: following the learner over time

A good tutor remembers. Knowledge-tracing research models the learner's state of knowledge to adapt what comes next. Recent approaches adjust these models to shifts in ability without retraining ^[5], or pair frozen LLMs with structured pedagogical memory for interpretable, state-of-the-art tracing ^[6]. For agents, memory based on an auto-updated knowledge graph offers robust long-term personalization ^[7]. Applied to review scheduling, content-aware retrieval improves learning efficiency ^[8].

• RAG: cuts hallucinations by grounding answers in a controlled corpus.
• External memory: personalises without degrading the base model.
• Knowledge tracing: adapts the path to the real state of knowledge.
• Limit: quality depends on retrieval and corpus, not just the model.

A reliable tutor is not the one that knows everything, but the one that knows where to look and remembers whom it is helping.

Learnya synthesis

These architectures have an extra privacy advantage: keeping learner memory outside the model, in a controlled space, eases sovereign data governance — an asset in European and Swiss contexts. Ground and remember, yes; but within a framework where you know exactly which data serves what.

When RAG really helps — and when it adds nothing

RAG is not equally useful everywhere. It shines when the answer must be grounded in a precise, authoritative corpus: the official curriculum, the institution's textbook, the teacher's course notes. In these cases, grounding generation reduces hallucinations and allows citing the source, which is pedagogically valuable. It adds far less on general knowledge the model already holds, as shown by the study where a RAG tutor did not beat a model without retrieval ^[4].

Quality then depends on an often-neglected link: retrieval itself. If the system brings back incomplete, outdated or off-topic passages, the answer inherits it — hence the documented cases of out-of-context answers despite RAG ^[2]. Investing in corpus structuring, document chunking and search relevance matters as much as the choice of generative model. Good RAG is first of all a good information-retrieval problem.

Learner memory adds a dimension: it enables personalised follow-up without rewriting the model. Keeping that memory external — profiles, knowledge graphs, histories ^[7] — avoids degrading the base model and offers a major privacy advantage. You know precisely what data is stored, where, and for what; you can audit, correct or erase it. Externalised grounding and memory thus form the foundation of educational AI that is both reliable and sovereign.

Toward grounded, accountable tutors

Combining RAG and memory sketches a credible architecture for tutors that are both useful and trustworthy: grounded in verifiable knowledge, able to remember the learner, and transparent about their sources. It is a concrete answer to the two most frequent criticisms of AI in education — hallucinating and forgetting.

The challenge of the coming years will be to industrialise these approaches without losing rigour: curating corpus quality, honestly measuring the residual error rate, and keeping learner data in a controlled space. A tutor that cites its sources and explains where its answer comes from is not only more reliable; it also teaches, by example, an essential skill: knowing what you rely on to assert something.