Continual Learning
Training AI models to accumulate new knowledge over time without overwriting what they already know.
Continual learning is the challenge of updating a model on new tasks or information without degrading performance on old ones. Standard fine-tuning is the problem: train on new data and the model tends to forget what it learned before. This is called catastrophic forgetting.
Why it matters: Deployed models go stale. The world changes, new products launch, regulations shift, user behaviour evolves. Running a full retrain every time is expensive and slow. Continual learning research is looking for ways to keep models current with targeted updates instead.
The core techniques:
- Sparse updates: only modify the weights most relevant to the new task, leaving others intact.
- Memory replay: interleave new training data with samples from previous tasks so the model stays sharp on both.
- Skill isolation: partition model capacity so new capabilities do not overwrite old ones (related to mixture-of-experts approaches).
Where research is heading: Lifelong learning for LLM-based agents is an active research area. Recent work explores how agents can accumulate skills across deployments rather than resetting with each update, through a mix of parameter-level updates and external memory.
The practical tension: Continual learning and RAG solve overlapping problems from different angles. RAG keeps knowledge current by retrieval at inference time. Continual learning bakes it into the weights. For enterprise teams, the choice depends on how structured the new knowledge is and how much latency you can afford.