SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent
The article introduces State-Adaptive Memory (SAM), a framework designed for long-horizon reasoning in artificial intelligence. SAM addresses the challenge of accessing relevant information from extensive interaction histories by consolidating ongoing interactions into compact memory cues. The framework has shown to outperform existing methods across various benchmarks, indicating its effectiveness in enhancing agentic reasoning capabilities.
- ▪SAM is a standalone framework that consolidates ongoing interaction into compact memory cues.
- ▪It allows agents to reconstruct temporally distant information according to their current needs without retraining the underlying model.
- ▪The framework has been optimized through expert-guided supervision and reinforcement learning.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.24468 (cs) [Submitted on 23 May 2026] Title:SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent Authors:Yuyang Hu, Hongjin Qian, Shuting Wang, Jiongnan Liu, Ziliang Zhao, Jiejun Tan, Zheng Liu, Zhicheng Dou View a PDF of the paper titled SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent, by Yuyang Hu and 7 other authors View PDF HTML (experimental) Abstract:Long-horizon agentic reasoning requires large language models to act over long interaction histories containing thoughts, tool calls, observations, and partial conclusions. The challenge is not merely that these histories grow long, but that information needed for the current decision may be scattered across distant steps and only become relevant later.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
