MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents
MemPilot lets an agent decide at runtime how much memory work is worth doing for a given query.
The paper frames the problem as a trade-off between accuracy, preprocessing cost, and latency in multimodal agent memory. MemPilot uses a reinforcement-learned policy to choose between existing query-agnostic memory and fresh query-specific curation from raw text and visual history. That policy also controls evidence volume, curation instructions, model choice, and visual access. The authors report better performance-cost-latency frontiers on five multimodal agent-memory benchmarks than existing trade-off-aware baselines. ArXiv · AI/CL/LG's note
The paper frames the problem as a trade-off between accuracy, preprocessing cost, and latency in multimodal agent memory. MemPilot uses a reinforcement-learned policy to choose between existing query-agnostic memory and fresh query-specific curation from raw text and visual history. That policy also controls evidence volume, curation instructions, model choice, and visual access. The authors report better performance-cost-latency frontiers on five multimodal agent-memory benchmarks than existing trade-off-aware baselines. ArXiv · AI/CL/LG's note
score 5