Research
Meta FAIR's Self-Improving Memory Lets Agents Discover and Distill Their Own Memory Prompts
Meta FAIR, with University of Washington and Johns Hopkins researchers, introduced Self-Improving Memory, a two-stage method that has an agent rewrite its LOAD/OFFLOAD memory prompts and then distills the resulting behavior into model weights via LoRA. Using Kimi-K2.6 as both task and meta agent, optimized harnesses averaged 56.0 on three Continual Learning Bench tasks versus 44.6 for the best human-written prompts. After distillation, the model averaged 43.1 with generic prompts versus 37.6 for SFT on generic trajectories, though the research code was not released.
Meta FAIRAI agent memoryContinual Learning Bench