New framework enables LLM agents to autonomously manage memory for extended reasoning tasks

Large language models deployed as autonomous agents struggle with complex, multi-step tasks that exceed their fixed context window limitations. A new research paper from arXiv addresses this challenge by proposing Agentic Memory (AgeMem), a unified framework that integrates long-term and short-term memory management directly into the agent's decision-making process.

The Memory Challenge#

Conventional approaches to memory in LLM agents treat long-term memory (LTM) and short-term memory (STM) as separate systems, often relying on fixed heuristics or auxiliary controllers to determine what gets stored and when. These rigid designs limit how flexibly agents can adapt to different task demands and prevent end-to-end optimization of the entire system.

How Agentic Memory Works#

Rather than hard-coding memory rules, AgeMem exposes memory operations as tool-based actions that the agent can invoke directly. The language model agent can then autonomously learn what information to store, retrieve, update, summarize, or discard based on the specific task at hand. This approach treats memory management as part of the agent's policy rather than as a separate component.

Training the Unified System#

To train agents to make effective memory decisions, the researchers introduced a three-stage progressive reinforcement learning strategy and developed a step-wise GRPO (Generalized Reward Preference Optimization) algorithm designed to handle the sparse and discontinuous reward signals that memory operations generate—a technical challenge that existing methods struggle with.

Demonstrated Results#

Testing across five long-horizon benchmarks showed that AgeMem consistently outperformed comparable memory-augmented baselines across different LLM backbones. The framework achieved measurable improvements in task performance, generated higher-quality stored memories, and made more efficient use of the available context window.