Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks (arxiv.org)

A recent study introduced "Memory-as-Action" (MemAct), a framework designed to enhance the performance of Long-context Large Language Models (LLMs) in executing long-horizon tasks. Traditional memory management methods often rely on external resources, failing to account for an agent's reasoning state, thus leading to less effective decisions. MemAct shifts this paradigm by integrating memory management into the learning process, allowing models to perform in-place editing actions—such as deletion and insertion—to refine their context dynamically. This end-to-end reinforcement learning approach results in optimized information retention and improved task execution. The significance of MemAct lies in its ability to dramatically boost efficiency: experiments demonstrate that the MemAct-RL-14B model achieves accuracy on par with models 16 times its size while simultaneously decreasing context length by 51%. The introduction of Dynamic Context Policy Optimization further mitigates training inefficiencies, preserving reasoning integrity amidst continuous context updates. This innovative framework represents a substantial advancement for the AI/ML community, promising to enhance the capabilities of LLMs by enabling them to adaptively manage context and optimize performance across diverse and complex tasks.