--- title: "Memory for learning and development assistants" description: "Memory for learning and development: how an assistant inside an LMS or a training product keeps a dated record of what a person has learned, stops a corrected mistake resurfacing, and traces guidance to the sessions behind it." canonical: https://past.dev/use-cases/learning last-updated: 2026-09-01 --- # Memory for learning and development Source: https://past.dev/use-cases/learning An assistant inside a learning product accumulates months of evidence about how one person learns and usually cannot use any of it after the session ends. Memory for learning and development has to keep each session dated, stop a mistake the learner has already corrected from resurfacing as current, and trace a recommendation to the sessions it came from. This page describes what that takes. ## What you are building An assistant that knows a learner rather than a question. Inside a learning management system it should open a session already aware that this person failed the same compliance module twice, corrected a misunderstanding about the approval threshold in March, and works through worked examples faster than through written guidance. The same record serves corporate learning and development, onboarding, compliance training, skills programmes and education products. What changes is the subject rather than the shape: in every case somebody is learning over months, and the evidence is spread across sessions nobody revisits. ## Why memory decides whether it works - **Every session starts from zero.** Months of evidence about how a person learns exists, and almost none of it is available once the session ends. - **A corrected mistake returns.** The long transcript where they got it wrong outranks the short exchange where they got it right, so the assistant coaches a misunderstanding the learner has already fixed. - **Guidance has no shape.** Advice built from the last answer alone cannot report that this is the fourth attempt at the same module, which is the thing a manager or an instructor needs to hear. - **Nobody can check a recommendation.** Guidance about a person's competence, offered without the sessions behind it, is guidance an organization cannot act on. ## What past.dev provides - **A dated timeline.** Each session goes in with its date, so what somebody knew in March and what they know now are separate, answerable questions. - **Superseded facts stay superseded.** A corrected mistake keeps its date in the history and does not return as the current state. - **Guidance you can trace**, so a manager or an instructor can read the sessions a recommendation was built from. - **One learner across sources**, so a session transcript, a submitted assessment and a manager's note resolve onto one person. - **Audience filtering**, applied during the search, so a shared assistant does not reach another team's or another cohort's material. ## Cost and accuracy A learner record in month nine costs about the same to query as in week one, because recall is not metered and the read per question is roughly fixed. That matters more here than in most places: an assistant priced per question is one a training team turns off once a programme scales past a pilot cohort. On accuracy, the property that decides it is temporal ordering. An assistant that resurfaces a corrected mistake teaches the mistake again, and in compliance training it can record somebody as not competent when the evidence says otherwise. See [benchmarks](/benchmarks) for the measured results and [pricing](/pricing) for the credit rules. ## Getting there 1. Send each session transcript, submitted assessment and instructor or manager note to `POST /api/v1/ingest` with the date it happened. 2. Backfill the history you already hold. Timestamps are backdatable, so previous programmes keep their original dates in one pass. 3. Call `POST /api/v1/recall` before generating guidance, and pass the returned evidence to the model. 4. Run the sync nightly. Ingestion is idempotent by content hash, so re-sending unchanged content is correct and free. ## Frequently asked questions ### Does this only suit education, or corporate training too? Both. The record is the same shape in either case: dated sessions about one learner, accumulated over months. Corporate learning and development, onboarding, compliance training and education products all read from it the same way. ### How does the assistant avoid repeating a correction the learner already made? The corrected value is the one supported by the most recent evidence. The earlier mistake keeps its date in the history rather than competing with the current state. ### Can a manager see why the assistant recommended something? Yes. Recall returns the sessions and the dated evidence behind an answer, so the basis for a recommendation is checkable rather than opaque. ## Related - [Productivity](https://past.dev/use-cases/productivity) - [Knowledge base](https://past.dev/use-cases/knowledge-base) - [Temporal knowledge graph](https://past.dev/glossary/temporal-knowledge-graph) - [Benchmarks](https://past.dev/benchmarks) - [Quickstart](https://past.dev/docs/memory-api/quickstart)