MemPrism Enhances AI Agent Memory for Long Tasks

This research introduces MemPrism, a novel memory framework designed to enhance the capabilities of long-horizon AI agents. A common challenge for these agents is the "representation mismatch," where relevant historical information is stored but not organized in a way that is immediately useful for the current decision-making process. Existing memory systems often rely on fixed representations, which can become inefficient and ineffective as task complexity and trajectory length increase. MemPrism addresses this by separating the long-term storage of experiences (an event stream) from the dynamic, task-conditioned construction of working memory. It allows agents to generate "relational views" of their past interactions, tailored specifically to the current task context. A lightweight view policy determines the structure, evidence range, outcome conditions, and granularity of these views. A deterministic composer then transforms these historical facts into a temporary, optimized working-memory view for the agent's task policy. Experiments on embodied and web-agent benchmarks demonstrate that MemPrism consistently improves task performance, particularly for longer and more complex trajectories. Crucially, it also reduces memory token consumption, making it more efficient. The learned view policy's ability to transfer across different Vision-Language Models (VLMs) without additional adaptation highlights its effectiveness as a general memory interface for AI agents.