Ebbinghaus and Forgetting

How does one study forgetting?

Is there a way to measure how forgetful we are?

That sounds hard. You cannot look inside a mind and watch a memory fade.

Hermann Ebbinghaus, a German scientist, found a simple way in. He treated forgetting as something you could test. In the 1880s, Ebbinghaus studied his own memory with lists of nonsense syllables. He wanted words that did not already mean much. A word like "table" brings images and associations with it. A made-up syllable carries less baggage, so it gives a cleaner test.

He learned a list, waited, and learned it again. If the second pass took less work, the first pass had left something behind. He called that leftover trace savings.

The savings score is commonly written as:

\text{Savings} = 100 \times \frac{ \text{original effort} - \text{relearning effort} }{ \text{original effort} }

If a list first took 100 repetitions to learn and later took 40 repetitions to relearn, the savings score would be 60 percent. The memory was harder to reach, but it was not gone.

The forgetting curve

Ebbinghaus found a pattern that still feels familiar. Forgetting is steep at first and slower later. A new memory can fade quickly after learning. After that early drop, the curve tends to flatten.

That pattern became known as the forgetting curve. It is not a universal law with one fixed shape for every person, task, or memory. Meaning, attention, sleep, practice, emotion, interference, and later reminders all matter.

The simple lesson is worth keeping: time changes memory. A system that treats an old memory and a recent memory as equal is missing something basic.

Forgetting is not always failure

Forgetting sounds like loss. But ask what would happen if you forgot nothing.

Old details would crowd out fresh ones. Every exception would stay active. Every draft, preference, and half-true note would keep trying to help. Forgetting reduces clutter. It lets stale details fade. It keeps yesterday's noise from steering today's decision.

This matters for AI memory because a system that remembers everything can get worse over time. It may retrieve old preferences, stale exceptions, irrelevant drafts, or facts that were true only for one moment.

Good memory needs decay. It also needs reinforcement. If a memory keeps helping, it should become easier to retrieve. If it stops helping, conflicts with newer evidence, or belongs to finished work, it should fade.

The bridge to ACT-R

Ebbinghaus was not proposing ACT-R. His work came much earlier. But he gives us a useful starting point: a memory can become easier or harder to retrieve over time.

ACT-R uses a more detailed model. Declarative memories are chunks. A chunk's availability depends partly on its history of use. Recent and repeated use can raise activation. Time can lower it. Context can help too.

That is why Ebbinghaus belongs near the start of this booklet. He gives us the older experiment behind the idea: memory is not only stored or missing. It can be easier or harder to bring back.

What this means for AI memory

For AI systems, Ebbinghaus is a warning against flat storage. Saving a record is not enough. The system also has to decide how much that record should matter later.

A useful memory layer should ask simple questions:

How recent is this memory?
Has it been reused successfully?
Does it still match the current context?
Has a newer decision replaced it?
Should it fade, stay available, or become core memory?

Those questions do not make an AI system human. They make the engineering more honest. Memory is not a pile of saved facts. It is a way for the past to matter, or stop mattering, at the right time.

Sources

Go back to Introduction to ACT-R, continue with ACT-R Memory vs Agent Memory, or return to the Concepts hub.