AI agents that browse, code, and reason across many steps have a memory problem — and a new paper proposes a fix.
Researchers introduced ECHO, a framework for agents that must juggle long sequences of tool use and decision-making inside fixed-size context windows. The core idea: instead of discarding old turns or squashing them into vague summaries, ECHO compresses each completed turn into a compact record and keeps track of which records actually contributed to a correct answer. When reinforcement learning kicks in, the system can route credit back through those source-indexed records to the specific evidence and choices that worked. On the BrowseComp-Plus benchmark, ECHO hit 43.4% held-out accuracy — clearing the GRPO baseline at 28.9% and the rolling-summary method SUPO at 36.1% — while using fewer turns and less total trajectory data.
This matters because the dominant workarounds for long-context agents — truncation and summarization — quietly sabotage the learning signal. If the model can't see which early observation led to a late correct answer, outcome-based training has no clean path to reinforce the right behavior. ECHO's source-indexed approach keeps that path open. The gains also held on zero-shot tasks across question answering, code generation, and information retrieval, suggesting the benefit isn't benchmark-specific.
The results are promising on paper, but academic benchmarks and production agent workloads rarely share the same pain points — so the real test will be whether the indexed-memory overhead stays manageable when task horizons stretch into the hundreds of turns.