AchiralAchiral

Concepts · Humans and machines

What Ebbinghaus measured through relearning, what savings means, and how AI builders can distinguish retrieval decay from deleting a record.

Published2026-07-080 reads

Review pending. Report issues to editor@achiral.ai.

Watercolor of a person clearing grass from worn stepping stones along a partly overgrown path leading to an old wooden gate.
The path needs work again, but the earlier work still helps. Recovering an existing route illustrates savings in relearning rather than starting from scratch.

Ebbinghaus Forgetting Curve

The Ebbinghaus forgetting curve describes how the benefit of earlier learning declines over time, measured by the effort saved when relearning material. In experiments published in 1885, Hermann Ebbinghaus found a steep early decline followed by a slower loss. His measure was savings in relearning, rather than a count of facts he could still recall.

The architecture overview separated stored knowledge from what is available to the current task. Ebbinghaus gives us an experimental starting point for that distinction: earlier learning can still help even when material is difficult to recall. For AI memory, this is a design analogy; a digital record does not fade merely because time passes.

Measure what earlier learning leaves behind

Ebbinghaus studied his own memory using lists of nonsense syllables. He chose them to reduce the influence of familiar meanings and associations. His account of the method explains how he prepared and learned the lists.

He learned a list, waited, and then relearned it to the same criterion. Comparing the effort required on the two occasions let him measure a benefit from the first learning even when immediate recall was difficult.

A common expression for the savings measure is:

Savings (%)=100×initial learning effort−relearning effortinitial learning effort\text{Savings (\%)} = 100 \times \frac{\text{initial learning effort} - \text{relearning effort}} {\text{initial learning effort}}

Both efforts must use the same measure and learning criterion. In a hypothetical repetition-count example, a list takes 100 repetitions to learn and 40 to relearn. Savings would be 60 percent: relearning required 60 percent fewer repetitions. It would not mean that the person could recall 60 percent of the list. Ebbinghaus's time-based calculations also included corrections described in his results.

Read the curve within its conditions

The familiar downward curve summarizes performance on a particular task under particular conditions. It does not give every person or kind of memory a fixed expiry schedule. Ebbinghaus himself cautioned against assuming that his numerical results applied generally.

Murre and Dros later replicated the savings experiment in another single-participant study. Their study supports the broad pattern while examining details that a smooth curve can hide, including variation across retention intervals and the possible role of sleep.

The distinction matters when interpreting a graph. Reduced savings shows that more work was needed to relearn. It does not directly reveal a physical memory trace being erased, or establish that elapsed time alone caused the change.

Connect forgetting to ACT-R retrieval

ACT-R represents declarative knowledge as chunks. In its base-level learning mechanism, a chunk's history of use contributes to activation. Recent and repeated use can make retrieval easier; the contribution of earlier uses declines with time. The ACT-R reference manual describes this mechanism alongside contextual influences and retrieval thresholds.

This gives a model a way to retain a chunk while making it harder to retrieve. A request can fail because activation falls below threshold, even though the chunk is still stored. ACT-R's retrieval account is more specific than simply applying an Ebbinghaus-shaped curve to every saved item.

Decide what should lose influence

For an AI system, lowering retrieval priority and deleting a record are separate operations. A builder might reduce the influence of an old preference while keeping its history. A deletion requirement, by contrast, cannot be satisfied merely by making the record unlikely to appear in context.

Consider the kinds of records in a shared work system. A temporary preference may stop being useful when a task ends. An old decision may remain essential when someone asks why a project changed direction. A newer decision may supersede the old one for current work while leaving both necessary to explain the history.

A decay policy needs to account for what the record represents, whether it has been superseded, and which task is asking for it. Repeated retrieval alone is a poor reason to strengthen a record: an outdated note could otherwise reinforce its own influence.

A useful review starts with a concrete question: when a decision is replaced, can the system favor the new decision for current work and still recover the old one for a historical question? Continue with Chunks and Declarative Memory for the representation and retrieval mechanisms behind that choice.


Sources

Go back to Introduction to ACT-R, continue with Chunks and Declarative Memory, or return to the Concepts hub.