SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent

Researchers introduced SAM, a state-adaptive memory framework for long-horizon AI agents.

SAM compresses interaction histories into compact cues while preserving detailed trajectories for later retrieval.

It is trained with expert-guided supervision and reinforcement learning to better adapt memory to task state.

Across benchmarks such as BrowseComp, BrowseComp-ZH, WideSearch, and HLE, SAM outperformed baseline methods.