Catastrophic Forgetting
Losing old skill while learning something new
- Catastrophic forgetting is an AI sharply losing what it used to do well while it's busy learning something new.
- The root cause is that there's no separate space for anything. Old knowledge and new knowledge share the exact same block of numbers.
- Training only grades against the data shown right now. Leave old data out entirely, and the old skill can collapse without any signal warning you.
- It gets held off by mixing in a bit of old data, pinning down important values so they can't drift, or attaching a separate new part instead.
- After teaching something new, you have to check the old scores alongside the new — the new score alone won't tell you.
Contents
1The analogy
Rub an iron rod with a magnet in one direction, over and over, and magnetism settles into the rod. It now pulls paper clips toward it, reliably enough to hold a note to a fridge door. The direction you rubbed it in has soaked into the whole bar, not just one end of it.
But start rubbing the same rod in the opposite direction, and as the new direction soaks in, the direction that was there before fades out. Before long, the rod responds only to the new direction. There's no separate compartment inside the rod to keep the old direction in. One rod, one direction — that's all it holds.
Catastrophic forgetting — an AI losing an old skill while it learns something new — works exactly like this iron rod.
2In detail
Old knowledge and new knowledge share the same space
A neural network is packed with adjustable numbers. Those numbers aren't split into a "cat compartment" and a "dog compartment." Both pieces of knowledge sit layered on top of the very same block of numbers.
Train it on a new task, and those numbers shift a little in whatever direction helps the new task. What happens to the old task in that new position never gets checked during training. The exact same thing happening to the rod as it gets rubbed the other way — the direction soaked in earlier getting pushed out — happens to these numbers too.
The lack of any dividing wall is the root of this. If there were separate compartments, new knowledge could just settle into an empty one. In reality, new knowledge reuses the very same numbers old knowledge was already using.
Only what's currently in view gets graded
Training measures how wrong something is and shifts the numbers by that much. But this wrongness only ever gets calculated on the data currently in hand. Feed in nothing but new data, and no matter how badly the old skill collapses, the number on screen keeps getting better. Training looks like it's going smoothly while the old ability is quietly draining out.
It drains fast, too. Old-task accuracy dropping below half after just a few passes isn't unusual. It doesn't fade gradually — it drops off a cliff, which is exactly why the word "catastrophic" is attached to the name.
It gets worse the more the new data differs in character from the old. Show a model that used to handle photos nothing but charts, or show a model that used to answer in a casual tone nothing but stiff formal documents, and the distance it has to travel is so far that the old spot gets shaken hard.
Three ways to keep the old direction from fading
The first is mixing in old data. Instead of feeding in only new data, a bit of the old material gets shown alongside it. When the original data isn't available anymore, the answers the old model used to give sometimes get mixed in as a substitute instead.
The second is pinning down important values. Numbers that contributed heavily to the old task get identified first, and a heavy penalty gets attached if they move too much. Less important numbers get left free to shift.
The third is attaching a separate part. The original numbers get left untouched, and a small part gets attached alongside them to handle just the new task. It's like leaving the rubbed rod exactly as it is and hanging a second rod next to it — and it means the add-on can be removed to fall back to the original state if needed.
Teach in small steps, and check as you go
Not moving too far in one go helps a lot on its own. Keep the step size small and the training short, and the numbers don't wander far from where they started. Mixing several tasks together in one pass, instead of swapping in a new task wholesale, also cuts down on how much gets pushed aside in the first place.
Above everything else, what matters is checking. Note the scores on a few old tasks before teaching something new, then measure those same tasks again once it's done. Look only at the new task's score, and you'll never find out what quietly disappeared.
It's worth getting into the habit of saving the state before teaching something new, too. If what got lost turns out to weigh more than what got gained, you can fall back to that saved state and try again with a different data mix or a smaller step size.
3More precisely
Catastrophic forgetting is the phenomenon where, across a sequence of tasks trained one after another, a later task overwrites the weights an earlier task relied on, causing the earlier task's performance to drop sharply. The field that studies this is called continual learning, and the three approaches above are formally known as rehearsal, regularization-based suppression, and structural separation.
The analogy breaks down in places, too. An iron rod can only hold one direction, but the numbers inside a neural network number in the billions, leaving genuine room for different abilities to coexist, layered on top of each other. Mix the data well while teaching, and several tasks can be held onto together, something no single rod could ever manage. There's one more gap: magnetism that leaves the rod is truly gone, but a weight isn't erased — it's shifted to a new position, and the old one is still recoverable in principle. That's why showing even a little of the old data again often brings the skill back far faster than learning it did the first time around, almost as if some trace of the old direction had lingered after all.
4Try it yourself
5Common misconceptions
It's easy to think a bigger model has room to remember everything, but actually size doesn't matter — showing it nothing but new data on repeat still drops the old task's score.
It's easy to think adding new knowledge always makes it smarter overall, but actually a task that was never re-checked can be quietly collapsing while the new score climbs.
It's easy to think this is similar to how people forget, but actually a person's forgetting fades gradually, while this phenomenon collapses sharply within just a few passes — the two work differently.
7One-line summary
In shortCatastrophic forgetting is new training overwriting the values that used to handle an old task, the way a rod only holds one direction at a time, and mixing in a little old data while checking old scores is the basic defense.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02