Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench
Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench
Meta FAIR, with University of Washington and Johns Hopkins University researchers, introduced Self-Improving Memory, a two-stage method in which a Kimi-K2.6 meta agent rewrites LOAD and OFFLOAD memory prompts and the resulting skills are distilled into the model's weights via a rank-64 LoRA adapter on attention projections. Self-found memory prompts averaged 56.0 across three Continual Learning Bench tasks versus 44.6 for the best human-written prompts, with Codebase reaching 64.3 (+20.6), while the distilled model averaged 43.1 versus 37.6 for SFT on generic trajectories. The post notes that 37 of 44 generated prompt revisions failed to beat the previous best, and that code was not released.
Source: MarkTechPost · marktechpost.com