SAM: State-Adaptive Memory Enhances Long-Horizon AI Reasoning
A new framework called State-Adaptive Memory (SAM) has been developed by researchers to enhance long-horizon reasoning in large language models (LLMs). This innovative method tackles the issue of LLMs managing extensive interaction histories, which encompass thoughts, tool calls, observations, and incomplete conclusions, often with pertinent information dispersed across various stages. Current strategies such as truncation, compression, or retrieval fail to adapt memory access according to the agent's changing state. SAM effectively compacts ongoing interactions into concise memory cues while retaining original trajectory pages for purpose-driven recall. Further details can be found in a paper published on arXiv (2605.24468).
Key facts
- SAM stands for State-Adaptive Memory.
- It is a standalone framework for long-horizon agentic reasoning.
- The method consolidates interaction history into compact memory cues.
- Raw trajectory pages are preserved for intent-driven recall.
- Existing approaches include truncation, compression, or retrieval.
- SAM explicitly models state-adaptive memory access.
- The paper is available on arXiv with ID 2605.24468.
- The research targets LLMs acting over long histories.
Entities
Institutions
- arXiv