Learning to Share: Selective Memory for Efficient Parallel Agentic Systems

The paper introduces "Learning to Share" (LTS), a learned shared-memory mechanism designed for parallel agentic systems, which allows for selective information reuse among multiple agent teams. LTS features a global memory bank and a reinforcement learning-based controller that optimizes memory usage to reduce computational overlap during parallel reasoning tasks. Experiments on AssistantBench and GAIA benchmarks indicate that LTS not only decreases runtime but also maintains or enhances task performance compared to traditional memory-free approaches, highlighting its potential for improving efficiency in multi-agent frameworks.