Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model

• Title: Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents (Jan 2026)
• Link: http://arxiv.org/abs/2601.01885v1
• Date: January 2026

Abstract

This paper proposes Agentic Memory (AgeMem), a unified framework that enables LLM agents to autonomously manage both long-term and short-term memory through learned tool-based actions. Moving beyond separate heuristic-based components, AgeMem integrates memory operations—such as adding, updating, retrieving, summarizing, and filtering—directly into the agent’s policy. To achieve this, the authors introduce a three-stage progressive reinforcement learning strategy and a step-wise Group Relative Policy Optimization (GRPO) method to handle sparse rewards. Experiments on five long-horizon benchmarks show that AgeMem significantly outperforms strong memory-augmented baselines.

Key Topics:

• Agentic Memory
• Large Language Model Agents
• Long-Term Memory (LTM)
• Short-Term Memory (STM)
• Reinforcement Learning (RL)
• Group Relative Policy Optimization (GRPO)
• Memory Management Tools
• Long-horizon Reasoning