FedAgentKE Shares Reasoning Abstractions Between Agent Frameworks Without Exchanging Raw Trajectories

arXiv 2607.21361·low signal

Weihao Li, Jun Bai and Ziyang Song target the fragmentation problem in memory-based agent systems: local experience stays trapped inside each framework, blocking cross-framework transfer. FedAgentKE applies iterative semantic knowledge distillation, aggregation, and adaptation so heterogeneous agent frameworks can collaboratively evolve transferable reasoning abstractions without sharing raw reasoning trajectories — a privacy property that matters for any org running agents over proprietary context. Experiments report consistent improvements in both cross-framework and cross-task settings, though the paper is light on absolute numbers.