A trained model is frozen in time — it knows the world as of its training cutoff and no further. Retraining from scratch to add new knowledge is enormously expensive. Continual learning is the pursuit of a better way: update models incrementally.
The core challenge: catastrophic forgetting
The obvious approach — just train the model on new data — has a nasty failure mode. Neural networks tend to overwrite old knowledge when learning new things, a problem called catastrophic forgetting. Teach the model this month's news and it may lose competence it had before. The whole field of continual learning is about adding new knowledge without erasing the old.
The hard part isn't learning something new. It's learning something new while remembering everything else.
Approaches
Several strategies help: protecting the weights most important to prior knowledge, replaying a mix of old and new data during updates, or adding new capacity (like adapters) for new knowledge while leaving the base frozen. RAG is a pragmatic sidestep — instead of updating weights, retrieve current information at query time.
Why it matters
Continual learning is the difference between models that need costly periodic retrains and models that stay fresh cheaply. For now, most systems handle "currency" through retrieval rather than weight updates, because retrieval is reliable and doesn't risk forgetting. But genuine continual learning — models that accumulate knowledge over their lifetime like people do — remains one of the field's important open problems.
