What is Continual Learning?

Definition

Continual learning is the training of one model across a sequence of tasks or data over time, so that it gains new abilities while keeping the ones it already has. Each round starts from the previous version of the model and updates it on new data. The central difficulty is catastrophic forgetting: an update that helps the new data can overwrite what earlier rounds taught the model. The research field dates back to McCloskey and Cohen's 1989 description of that interference.

For AI agents the term is used more broadly. LangChain's Harrison Chase splits continual learning for agents into three layers in "Continual learning for AI agents" (5 April 2026): the model, whose weights change; the harness, the code and built-in instructions that drive the agent; and the context, the memory and per-user instructions it is given. Harness and context changes are quick to make and easy to roll back. Continual learning at the model layer trains the next model version on the agent's own traffic, which changes what the model itself can do.

Product writing often says continuous learning for the same idea. Arize's developer guide (last reviewed 16 August 2026) notes that the two terms are used interchangeably in product discussions.

Why It Matters

A model trained once starts going stale the day it ships. Users change how they ask, tools and upstream data change shape, and new edge cases appear. Google Research's sleep framework post (27 October 2026) describes deployed LLMs as largely static, with a fixed knowledge cutoff. Continual learning is how a model keeps up with the work it is doing.

For an agent that repeats the same task at volume, the model layer also sets the cost. Instructions and examples that travel with every prompt can be trained into the model, so each task needs fewer tokens, and a specialist trained for the task usually costs less per token than a frontier model. Each training round is also a chance to lose something, which is why every new version needs an evaluation gate before it takes traffic.

How It Works

At the model layer, continual learning runs as a loop in the background while the current model keeps serving. The agent's requests, responses and outcomes are logged. Candidate training examples are drawn from that traffic and approved by the team that owns the agent. A new version trains on the approved examples, mixed with examples from earlier rounds (replay) so it keeps what those rounds taught it. The new version is scored against the team's evaluations: a task set that defines the quality bar and a regression set of cases the current model already handles well. It goes live only with the team's sign-off, and if it falls short, the current model stays in service.

Techniques that reduce forgetting between rounds include replay, regularisation such as elastic weight consolidation, parameter-efficient training such as LoRA, and orthogonal subspace methods such as O-LoRA and OSFT. Each one lowers the risk, and every version is still measured before release.

Example Use Case

A support agent classifies and drafts replies for incoming tickets. After a product launch, a new category of question appears that the model routes badly. The team approves a few hundred corrected examples from the agent's traffic, and a new model version trains on them alongside examples from earlier rounds. The new version is scored on the task set and on a regression set covering the categories the current model already handles. It holds the bar on both, the team signs off, and it replaces the current model. Had it slipped on the regression set, the current model would have stayed in service and the failed cases would have gone into the next round.

Key Takeaways

• Continual learning trains one model over a sequence of rounds while keeping what earlier rounds taught it.
• For AI agents it happens at three layers: model, harness and context.
• At the model layer, the agent's own traffic trains each new model version.
• Catastrophic forgetting is the main risk, reduced by replay, regularisation, LoRA and orthogonal subspace methods.
• An evaluation gate with a regression set decides whether a new version replaces the current one.