Self-Improving Memory: Meta-Optimization and Harness Distillation for Continual Learning Agents | RAM @ Meta AI

facebookresearch.github.io14 min read

Researchers at Meta AI introduce Self-Improving Memory, a method that allows language model agents to automatically discover and optimize their own memory-management strategies through a Discover-Demonstrate-Distill pipeline. By combining harness optimization with harness distillation, the model internalizes discovered memory skills into its weights, outperforming both human-designed memory prompts and generic baselines across continual learning benchmarks. Open article →

More on this topic