ECHO: Prune to act, trace to learn with selective turn memory in agentic RL

2026-06-30 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

ECHO is a novel selective turn-memory framework designed for long-horizon language agents operating under bounded context windows. It addresses two key limitations in existing context-management methods: the progressive removal or collapse of historical observations, which hinders fine-grained evidence reuse, and the loss of an explicit path for aligning policy updates with successful outcomes once original turns are no longer source-addressable. ECHO tackles these issues by compressing each completed environment turn into a compact memory record, reconstructing bounded policy contexts by selectively choosing from these records, and reusing the selected source indices to route positive outcome credit. On the BrowseComp-Plus benchmark, ECHO achieves 43.4% held-out accuracy, significantly outperforming GRPO (28.9%) and the rolling-summary baseline SUPO (36.1%), while also requiring fewer turns and lower trajectory volume than SUPO. Furthermore, the trained policy demonstrates improved zero-shot generalization across multi-objective QA, code generation, and deep information-seeking benchmarks, utilizing both dense and MoE backbones.

Key takeaway

For Machine Learning Engineers developing long-horizon language agents, ECHO offers a robust solution to context window limitations and policy alignment challenges. You should consider implementing selective turn memory and source-indexed reconstruction to maintain fine-grained evidence and enable traceable learning. This approach can significantly improve agent performance, as demonstrated by ECHO's 43.4% accuracy on BrowseComp-Plus, and enhance zero-shot generalization across diverse tasks like QA and code generation.

Key insights

ECHO uses selective turn memory and source-indexed reconstruction to enable fine-grained evidence reuse and traceable learning in long-horizon language agents.

Principles

• Selective turn memory prevents history collapse.
• Source-indexed reconstruction enables traceable learning.
• Compact memory records improve context efficiency.

Method

ECHO compresses environment turns into memory records, reconstructs policy contexts by selecting records, and routes positive outcome credit via source indices to align policy updates with supporting evidence and selection actions.

In practice

• Apply selective memory for long-horizon tasks.
• Use source indexing for outcome-based RL.
• Improve generalization with trained policies.

Topics

• Agentic RL
• Context Management
• Language Agents
• Memory Networks
• Policy Alignment
• Zero-shot Generalization

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Drawing on Memory: Dual-Trace Encoding Improves Cross-Session Recall in LLM Agents — Dual-trace memory encoding enhances LLM agent recall by adding contextual "scene traces" to factual memories.
• MemPro: Agentic Memory Systems as Evolvable Programs — Evolving the entire memory construction-retrieval pipeline as a program, not just the memory bank, significantly enhances long-horizon autonomous agents.
• Learning What to Remember: Observability-Safe Memory Retention via Constrained Optimization for Long-Horizon Language Agents — Memory retention for long-horizon agents is a constrained stochastic optimization problem requiring observability-safe learning.
• Your AI Agent Isn’t Dumb. It Has ADHD — LLMs forget instructions mid-conversation due to context window attention bias, causing significant performance drops.
• Multi-Turn Reasoning When Context Arrives in Pieces: Scalable Sharding and Memory-Augmented RL — Training LLMs with compact rolling memory significantly improves multi-turn reasoning by mitigating "Lost in Conversation" degradation.
• Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity — Memora decouples memory content from retrieval, balancing abstraction and specificity for scalable AI agent memory, achieving leading performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.