Integrating unified memory management into a single agent without auxiliary expert models

Determine a method to integrate unified management of long-term memory and short-term memory directly within the policy of a single large language model agent, eliminating reliance on external expert models for memory control to enable end-to-end deployment with reduced inference cost and training complexity.

Background

The paper argues that many existing agent systems control memory using an auxiliary expert LLM, which increases inference costs and training complexity. Prior work typically treats long-term memory and short-term memory as separate components managed by heuristics or external controllers, leading to fragmented optimization and limited adaptability.

Within the discussion of challenges for achieving unified memory management, the authors highlight a deployment obstacle: removing dependence on external expert models while keeping memory control integrated inside the agent. They explicitly state that this integration problem remains open, motivating their proposed Agentic Memory framework as a step toward addressing it.