ECHO: Prune To Act, Trace To Learn With Selective Turn Memory In Agentic RL

arXiv cs.AI · 18 hours ago

ECHO is a selective turn-memory framework for language agents in reinforcement learning that compresses environment interactions into indexed records and reconstructs context windows while preserving traceability for credit assignment. On BrowseComp-Plus, ECHO achieved 43.4% accuracy compared to 28.9% for GRPO and 36.1% for the rolling-summary baseline SUPO. The approach enables policies to improve credit assignment to supporting evidence while using fewer turns and lower trajectory volume than competing methods.