MemPrism: Task-Conditioned Relational Memory Views for Long-Horizon Agents

2026-08-06

AI Summary

This work addresses the challenge that long-horizon agents often fail to effectively reuse stored experiences due to representational mismatches between past memories and current task contexts. To overcome this, the authors propose MemPrism, a framework that decouples persistent memory from task-time working memory and dynamically constructs relational memory views conditioned on the current task. MemPrism employs a lightweight view policy, a deterministic composer, and a renderer to generate task-adapted ephemeral memory structures from event streams, enabling seamless transfer across diverse vision-language models without fine-tuning. Experiments demonstrate that MemPrism significantly improves performance on long-horizon embodied and web-based agent benchmarks, particularly excelling in extended-task scenarios while substantially reducing memory token consumption.

Abstract

Long-horizon agents rely on memory to reuse experiences, yet existing memory systems often assume that evidence can be directly consumed through a fixed representation. This leads to representation mismatch, where relevant information is available but not organized for the current decision. To this end, we propose MemPrism, a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory. MemPrism records interactions as the event stream and dynamically constructs relational views according to the current task context. A lightweight view policy selects the relation structure, evidence range, outcome condition, and granularity, while a deterministic composer and render transform historical facts into a temporary optical working-memory view for a frozen task policy. Experiments on long-horizon embodied and web-agent benchmarks show that MemPrism consistently improves the task performance, especially as trajectories become longer, while reducing memory token consumption. Furthermore, the learned view policy transfers across different VLMs without additional adaptation, demonstrating the effectiveness of task-conditioned relational views as a general memory interface for agents.

Problem

• long-horizon agents
• memory systems
• representation mismatch
• relational memory
• task-conditioned

Innovation

• task-conditioned memory
• relational views
• long-horizon agents
• working memory
• memory efficiency