Meta FAIR proposes routed continual learning architecture Routed CL, API costs can be reduced by 65% to 72%

When facing similar tasks, having agents reuse the experience they have accumulated from one another is a natural idea. Unfortunately, there is no single way of retaining experience that works well in all scenarios, nor is it possible to know in advance which future problems a given piece of experience will help with.

For continuously running agent systems, past trajectories can indeed reduce repeated trial and error on subsequent tasks, but once the experience from different tasks is mixed together, it can also cause interference; if all historical trajectories are retained uniformly, the details are indeed the most complete, but this increases context and API costs; and if all experience is compressed into a single form of memory, information that is useful only for specific tasks may be lost.

To this end, the Meta FAIR team proposed "Routed Continual Learning" (Routed CL). This architecture sorts historical trajectories into different clusters and maintains a separate set of memories for each cluster. Before each run, an LLM-based router uses the current problem, descriptions of the available memory systems, and summaries of the existing clusters to first determine which type of historical experience this task is closest to, and finally decides whether to reuse a certain cluster or create a new one.

This architecture has two core features. The first is transparent deployment: the architecture can be plugged directly into existing agent systems, without requiring the agents to coordinate with one another and without changing how they originally operate; the second is routing: Routed CL not only decides which tasks should share experience, but also selects, from a fixed library of memory systems, how experience should be stored. Some tasks are suited to retaining raw trajectories, some are better suited to distilling reusable reasoning strategies, and others may require structured numerical states. Therefore, all tasks do not have to use the same memory representation.

Figure | Overview of Routed Continual Learning

They also designed a Context Memory Adapter, which is responsible for connecting the memory systems and the agent. Before a task begins, it writes the memories in the corresponding cluster into the agent's context, and after the task is completed, it updates that cluster with the current trajectory.

They evaluated this architecture on a task stream that mixed a variety of tasks. The evaluation included 301 instances from six CL-Bench tasks and 97 banking instances from τ³-Bench, each instance being a multi-turn customer service conversation. Instances from different sources were interleaved, and the system had to handle them without explicit instance labels. The experiments covered four language models: Gemini 3.7 Flash, GPT-5.6 Sol, Opus 4.7, and Opus 5, and each model was run 5 times to reduce the influence of randomness.

Figure | Five memory systems, grouped by the form in which they retain state

The results show that, compared with Global ICL, which uses only a single global context to learn from history, Routed CL improves the normalized reward by an amount equivalent to 15% to 30% of the maximum achievable normalized reward, while reducing the estimated API cost by 65% to 72%. Its average normalized reward is also higher than that of the ICL-clustering-only approach.

Figure | The relationship between average normalized reward and average estimated API cost over five paired runs

Across the four models, the average normalized reward of Routed CL was higher than that of the three comparison approaches: Stateless ICL, Global ICL, and Routed ICL. On this basis, they argue that the key to continual learning is not remembering as much as possible, but rather making two judgments at the same time: in what form experience should be retained, and which experiences are worth putting together for reuse.

Figure | Normalized reward under five paired full runs

The results also show that Global ICL varies greatly in performance across different evaluation settings. In the original CL-Bench, each type of task was run separately, and in that case Stateless ICL was one of the strongest approaches; but when multiple tasks were interleaved into the same stream, Global ICL increased the average normalized reward by only 0.015 compared with the stateless condition. This shows that the real problem with global memory is not just its limited capacity, but that the experience from different tasks interferes with one another.

Of course, this does not mean that Routed CL has solved all the problems of continual learning. They acknowledge that the current experiments process tasks in sequence, with each rollout using one language model as the actor, which simultaneously takes on all the model-mediated roles in the memory systems. Although the system's design has already taken into account isolation between different memory clusters and update consistency, it has not yet evaluated concurrent updates, reading stale versions of state, or different models reusing the state of the same cluster.

They also point out that in the future, a single task might be able to call multiple memory clusters at the same time; and a new run trajectory might also update multiple clusters at once, or even write to memory in different forms.

For more technical details, see the original paper.

English text is being prepared.