MemPrism: Task-Conditioned Relational Memory for Long-Horizon Agents
A new study on arXiv (2608.06745) introduces MemPrism, a framework designed to improve long-horizon agents by managing relational memory based on specific tasks. It separates persistent experiences from the working memory used during decision-making, addressing the issue where relevant information isn't readily structured for use. MemPrism records interactions as a stream of events and creates relational perspectives that adapt to whatever task is at hand. It employs a streamlined policy to define the relation structure and conditions, while a deterministic composer transforms past data into a temporary working-memory view for a fixed task policy. Testing shows that MemPrism significantly enhances task performance in long-horizon embodied and web-agent benchmarks. The research is categorized under artificial intelligence and machine learning and was recently shared on arXiv.
Key facts
- Paper ID: arXiv:2608.06745v1
- Announce Type: new
- Proposes MemPrism, a task-conditioned relational memory framework
- Separates persistent experience storage from decision-time working memory
- Addresses representation mismatch in memory systems
- Uses a lightweight view policy to select relation structure, evidence range, outcome condition, and granularity
- Experiments on long-horizon embodied and web-agent benchmarks show consistent improvement
- Published on arXiv
Entities
Institutions
- arXiv