Use casesEducation
Tutors that remember the learner.
An assistant that knows what a student covered in March, what they got wrong in June, and what changed between the two.
Maya got equivalent fractions right unprompted today, three weeks after struggling with them. She is still inverting the divisor when dividing.
- Equivalent fractions: secureReplaces the earlier value
struggling, 18 Julkept, dated - Division by a fraction: inverts the divisor
- Works faster from a worked example
What should the next session cover?
Division by a fraction. Equivalent fractions closed on 9 August, so there is no need to re-teach them.
The problem
Why the AI tutor never gets past the first lesson
01
A term of sessions, all starting from zero
A hundred sessions of evidence about how a student learns, and almost none of it survives the session boundary.
02
The misconception it already corrected, back again
The transcript where they got it wrong is longer and ranks higher than the short exchange where they got it right.
03
Feedback with no shape to it
Guidance built from the last answer alone cannot say the useful thing: that this is the fourth time division has gone wrong.
What memory changes
A learner record that accumulates
01
A timeline, not a snapshot
Each session goes in dated. What a student knew in March and what they know now are both answerable questions.
02
Superseded facts stay superseded
A misconception the student has since corrected does not resurface as current. It stays in the history with its date.
03
Answers you can trace
Every answer comes back with the sessions it drew on, so an instructor can check the basis for a recommendation.
Economics
The bill does not grow with the history.
The usual alternative is to send the whole record to the model on every question, so cost climbs with every message you have ever stored. We read a roughly fixed amount per question whether the record holds a hundred thousand tokens or ten million.
- 23x cheaper per question than sending 1M tokens of history, on the same model
- 232x cheaper at 10M tokens, where no production model can hold the context at all
- Around 96% less context read per question at 1M tokens of history
Integration
Three calls, and nothing to host.
One input shape: text plus a timestamp. You do not pick an embedding model, chunk anything, maintain entity tables or write prompts that decide which of two conflicting statements is current.
- Four endpoints and one bearer token, called directly over HTTP
- Ingestion is idempotent by content hash, so a crashed backfill just re-runs
- SOC 2 Type II, ISO 27001 and ISO 27701, with the API inside that same scope
Why Revo Memory
Why learning teams pick Revo Memory
Cost does not scale with the term
A student record in week thirty costs about the same to query as in week one, because the read per question is roughly fixed.
Backfill the whole year
Timestamps are backdatable, so an existing history loads onto its real dates in one pass.
Idempotent ingestion
Re-sending unchanged content costs nothing, so a nightly sync can be dumb and still be correct.
No pipeline to build
No chunking, no embedding model to choose, no vector index to operate.
Permissioned by audience
The retrieval filter is applied during the search, so a shared assistant cannot surface another cohort's material.
Four endpoints
Ingest, recall, answer and delete. The whole integration is a handful of HTTP calls.
More use cases
The same engine, other shapes of work.
Tell us what you are building.
Keys are issued by hand while the API is in early access. Bring your volumes and the shape of your integration, and we will tell you straight whether it fits yet.