Memory-Centric Continual Learning for AI Agents

Memory-Centric Continual Learning for AI Agents. By Dongfang Li and twenty-seven others. Published on the twenty-first of October. No code repository is linked to it.

Agents built on foundation models can reason, plan, and use tools. What they struggle with is learning from experience over a long deployment, because their weights are fixed and their context is limited. External memory systems try to close that gap. Mostly, though, they act as passive retrieval caches, and they are judged chiefly on whether they can recall stored facts or past interactions. Whether memory can drive genuine continual learning is, the authors say, still an open question. Recall tells you that something was kept. It does not tell you whether the agent behaves any better for having kept it.

Their proposal is called Memory-Centric Continual Learning, or MCCL, and it rests on one change of criterion. Memory should be judged by whether it improves future actions, not merely by what it retrieves. MCCL has three main parts.

The first is a three-axis taxonomy. It sorts memory by where state is stored, called substrates; by what form that state takes, called assets; and by how it is managed, called policies. This taxonomy defines a lifecycle with two loops. The inner loop updates memory state. The outer loop refines the management policies, based on task feedback.

The second part is an evaluation protocol. It measures three things: retention under interference, adaptation to change, and transfer across tasks. The point of measuring those is to separate genuine behavioural gains from what simple context expansion would give anyway. The third part is a memory scaling framework. It models how agent performance scales with experience, with storage capacity, and with compute budget, across both non-parametric and parametric memory.

One thread runs through all three parts: memory is treated as something that learns rather than something that stores. The taxonomy gives that learning a place to happen, in the outer loop that refines how memory is managed. The evaluation protocol is built to tell real learning apart from a larger context window. And the scaling framework asks the question one would ask of any learner, how performance grows as experience, storage, and compute grow, covering memory held outside the model's weights as well as inside them.

The contribution is a way of organising and judging agent memory rather than a single new memory system. The reframing is from passive lookup to continuous learning. Recall of stored facts stops being the test; better future action becomes it. The authors present the three components together as a principled route toward self-improving autonomous agents.