Routed Continual Learning
An agent experience management architecture that reframes continual learning as a routing problem. It clusters historical trajectories and uses an LLM router to select memory systems on demand, balancing the preservation of detail against control of context cost.
Timeline (last 90 days)
Routed Continual Learning (Routed CL) is a continual learning architecture that decides how to partition and retain experience, with its core idea being to treat continual learning as a routing problem—selecting not only the learning mechanism but also the state cluster, rather than committing globally to a single representation.
Routed CL groups past trajectories into multiple clusters, each maintained by one memory system from a fixed library of memory systems; before each run, an LLM-based router decides, based on the problem statement, memory system descriptions, and cluster summaries, whether to reuse an existing cluster or create a new one.
The Routed CL architecture introduces a Context Memory Adapter, which mediates between the memory system and the agent loop, converting the memory system's output into context information usable by the agent and ensuring that different memory mechanisms can all be seamlessly integrated into the agent's decision process.
The research team evaluated Routed CL on a heterogeneous task stream interleaving 301 instances from six CL-Bench tasks and 97 banking-domain instances from τ³-Bench; the experiments covered four language models and five matched rollouts were performed.
Compared with a single global in-context learning (ICL) history, Routed Continual Learning improves normalized reward by 15-30% of the maximum achievable normalized reward.
Relative to Global ICL, Routed Continual Learning reduces estimated API cost by 65-72%.
The Routed CL method depends on the judgment accuracy of the LLM router and the design quality of the memory system library.
How to automatically expand the memory system library and how to handle transfer learning between clusters remain directions worth exploring in the future.
All Facts (8)
Routed Continual Learning (Routed CL) is a continual learning architecture that decides how to partition and retain experience, with its core idea being to treat continual learning as a routing problem—selecting not only the learning mechanism but also the state cluster, rather than committing globally to a single representation.
50%UnverifiedRouted CL groups past trajectories into multiple clusters, each maintained by one memory system from a fixed library of memory systems; before each run, an LLM-based router decides, based on the problem statement, memory system descriptions, and cluster summaries, whether to reuse an existing cluster or create a new one.
50%UnverifiedThe Routed CL architecture introduces a Context Memory Adapter, which mediates between the memory system and the agent loop, converting the memory system's output into context information usable by the agent and ensuring that different memory mechanisms can all be seamlessly integrated into the agent's decision process.
50%UnverifiedThe research team evaluated Routed CL on a heterogeneous task stream interleaving 301 instances from six CL-Bench tasks and 97 banking-domain instances from τ³-Bench; the experiments covered four language models and five matched rollouts were performed.
50%UnverifiedCompared with a single global in-context learning (ICL) history, Routed Continual Learning improves normalized reward by 15-30% of the maximum achievable normalized reward.
50%UnverifiedRelative to Global ICL, Routed Continual Learning reduces estimated API cost by 65-72%.
50%UnverifiedThe Routed CL method depends on the judgment accuracy of the LLM router and the design quality of the memory system library.
50%UnverifiedHow to automatically expand the memory system library and how to handle transfer learning between clusters remain directions worth exploring in the future.
50%