Meta FAIR's Routed Continual Learning lifts agent memory gains 15-30% on mixed task streams, cuts API cost 72%

dair_ai · x · 2026-10-31

Meta FAIR paper: global memory barely helps interleaved agent tasks

Meta FAIR (incl. Ilya Mironov) published Routed Continual Learning, addressing how agents learn from past runs:

Problem: the usual single global history gains almost nothing over no memory when tasks are interleaved (+0.015 normalized reward on average).
Method: past trajectories are clustered, each cluster maintaining state in one memory system from a fixed library; before each run an LLM router reuses a cluster or creates a new one, with a Context Memory Adapter injecting memory outputs into the agent's context.
Results: on a stream of 301 CL-Bench instances mixed with 97 τ³-Bench banking instances across four models, Routed CL raised reward by 15–30% of the maximum attainable reward over global history and cut estimated API cost by 65–72%.
Limitation: runs were sequential; parallel agents updating one cluster remain untested.

More from coding & agent