Meta FAIR proposes Routed Continual Learning to split agent memory into clusters
FAIR at Meta proposes Routed Continual Learning, which groups past agent runs into clusters with their own memory systems and cuts estimated API cost by 65% to 72%.
FAIR at Meta has proposed Routed Continual Learning, an architecture that stops agents from sharing one growing pool of past experience. Instead, it sorts that experience into clusters, each with its own memory system. The main evidence is a mixed stream of 301 CL-Bench instances and 97 banking instances from τ³-Bench. In that stream, a single global history gained only +0.015 normalized reward over running with no memory at all. The authors report that the routed design raises reward over that global history by 15% to 30% of the maximum attainable normalized reward across four language models. It also cuts estimated API cost by 65% to 72%.
The design has two parts. Past trajectories are grouped into clusters, and each cluster keeps its state in one memory system drawn from a fixed library. Examples include retained trajectories or reusable reasoning strategies. Before each run, an LLM router decides whether to reuse an existing cluster or create a new one. It bases that choice on the problem statement, the memory system descriptions and cluster summaries. A second component, the Context Memory Adapter, takes the memory system's output and adds it to the agent's initial context. Then it writes the new trajectory back into the chosen cluster. The adapter relies on hooks that agent frameworks already provide.
The routing idea addresses a tradeoff the paper describes at the outset. A global history keeps every detail but gets expensive, and it mixes together experience that may not belong together. Compressing it into one fixed representation limits the cost but can discard valuable details. Routed CL tries to avoid committing to one representation globally. Instead, the router picks both the memory mechanism and the cluster for each run. The abstract's conclusion, in its own words, is that continual learning can be approached "as a routing problem."
The gains are measured against a specific baseline, and the numbers need that context. The 15% to 30% figure is a fraction of the maximum attainable normalized reward, not a percentage gain on raw reward. The abstract gives it as a range across four models and five matched rollouts. The 65% to 72% cost reduction is an estimate, measured relative to the global in-context learning history. The paper also reports that the routed design achieves higher mean normalized reward than clustering that uses only in-context learning.
The evaluation has limits the authors state themselves. Runs are sequential, so the paper does not test several agents updating the same cluster at once. The DAIR.AI Academy summary of the paper adds that the authors do not evaluate simultaneous updates, stale reads or cross-model reuse of cluster state. The results therefore cover sequential runs only.
What changed
FAIR at Meta proposes Routed Continual Learning, which sorts past agent trajectories into clusters that each keep their own memory system instead of one global history.
Type: Research
Why it matters
A single shared memory for many agent tasks may add little while raising costs. Teams running varied agent workloads may gain from routing memory by task type.
What you can do
Compare one global agent memory against task-clustered memory on your own workloads before scaling.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works