Ebbinghaus’s Forgetting Curve and AI Coaching: How Blink AI Operationalises It

In 1885 Hermann Ebbinghaus measured something hard: you forget about half of what you learn within an hour, two-thirds within a day, three-quarters within a month. 140 years on, corporate training is still not designed against his curve. The training finishes, the box gets ticked, and three weeks later nobody can put the knowledge to work in the field.

This post is about why the curve is still valid, why classical spaced-repetition tools fail in a corporate flow, and how Blink AI's three-layer architecture bends the curve back. The longer theoretical version of this argument was published earlier as a LinkedIn article; this one is what that framework looks like inside a real product.

The curve hasn't moved

Ebbinghaus used himself as the subject, memorising nonsense syllables, and recorded what fraction he could recall after specific intervals. The chart below sketches his classical curve:

Later research showed the numbers shift with meaningful content, motivation and sleep quality. The underlying behaviour did not change: learning decays exponentially unless re-encountered. Corporate data tells the same story. A short recall test one week after a safety module typically lands in the 30-40% band — the training looked complete, but the knowledge slid out under the feet.

Why spaced repetition works

The fix has been known since the 19th century: see the same content again at increasingly long intervals, and the curve flattens. Three mechanisms drive it:

Why classical spaced repetition stalls in a corporate flow

SuperMemo, Anki and Leitner have served individual learners well for decades. Move them into a corporate context and they hit three walls:

• Static scheduling. Classical algorithms (SM-2 and descendants) adjust the next interval only on the last correct/incorrect signal. They ignore concept difficulty, user's prior performance on neighbouring concepts, and the content type.
• Format mismatch. The card format (front / back) is too thin for corporate learning. Leadership, sales and safety concepts don't fit on a card — they need scenarios, dialogue, role-play.
• Responsibility transfer. Classical tools say "reviewing is your job". Corporate learners don't carry that intrinsic motivation; the training flow itself has to absorb the repetition.

The three layers of Blink AI

We designed Blink AI around three cognitive phases — encoding, storage, retrieval — with an AI component on each one, all stitched into the learner's flow.

Layer 1 — Encoding: Socratic dialogue + semantic context

Deeper processing means slower forgetting (levels-of-processing). On Blink the learner is never a passive viewer; through GROW and RPM frames the system has them construct their own answer. When a topic opens, the AI asks first: "Have you faced this before?" — then connects the answer to the content. A RAG layer pulls pieces from the content base that are semantically related to the learner's own prior statements, so every concept lands attached to a node they already know.

Layer 2 — Storage: gamified micro-learning

Long lecture videos are replaced by 10-15 minute modules. Each module is a short scenario, a short role-play and a closing "what did you take from this?" synthesis. That synthesis is not a quiz: the AI grades not for right/wrong but for which layer of the concept the learner reached. The output becomes a competency signal in the database: understood / needs_review / mastered.

Layer 3 — Retrieval: AI-driven spaced repetition

This is the layer that bends the curve. Unlike SM-2, Blink computes the next repetition timing against four inputs:

• The concept's history of understood / needs_review / mastered transitions.
• A difficulty index for the concept — tagged at content creation, auto-adjusted from cohort performance.
• State of semantically related concepts — when one neighbour is at the edge of forgetting, surfacing them together strengthens retrieval.
• Behaviour signals — app-open frequency in the last week, reply latency, "aha" markers in the dialogue.

The output isn't a push notification — it's the next micro-module in the learner's flow simply being the right one. The rhythm runs on the system's knowledge, not the learner's discipline.

What we measure

You cannot judge this architecture from completion rates — completion is only the starting point of the curve. The signals we actually track:

• 30-day recall score: accuracy on synthesis prompts a month later.
• Mastery transition: ratio of concepts moving from needs_review to mastered.
• Review-reason signal: when the learner asks for a repeat, why (struggled, curious, "aha"). The AI reads these as direct product signal.
• Behaviour transfer: real-world decisions after training — only possible when the customer connects their downstream systems. xAPI / native API surfaces exist for exactly this kind of integration.

Limits and what's next

This architecture is not magic. Content needs to be written for Socratic prompting — uploading old PDFs and expecting the same outcome won't work. Spaced-repetition efficiency depends on the learner adopting the app and producing enough behaviour signal; the first 2-3 weeks are a cold-start window. And feeding the retrieval layer with real field signal needs an integration pipeline on the customer side — usually a separate project in most organisations.

Next on our side: cohort-driven content difficulty auto-calibration, an LRS export path (xAPI), and richer multi-modal behaviour signal (email opens, calendar, real meeting outputs).

Closing

Ebbinghaus didn't just draw a graph 140 years ago — he left a question modern corporate learning still has to answer: "what's still there on day 30?" The answer to that isn't content alone; it's the coaching, gamification and retrieval that you wrap around the content over time. Blink AI is what those three look like when you commit to stitching them together.