Meta FAIR's Self-Improving Memory scores 56.0 on Continual Learning Bench
Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench
Meta FAIR, with researchers from the University of Washington and Johns Hopkins University, introduced Self-Improving Memory, a two-stage method that lets a Kimi-K2.6 agent rewrite its own LOAD and OFFLOAD memory prompts and then distill the resulting skills into LoRA weights. The optimized harness averaged 56.0 across the Database, Codebase and Spectrum tasks of Continual Learning Bench, versus 44.6 for the best human-written prompts, with Codebase gaining 20.6 points to 64.3. After distillation the model runs the generic prompts and still averages 43.1, against 37.6 for SFT on generic trajectories, though 37 of 44 meta-agent revisions failed to beat the previous best.
Source:MarkTechPost · marktechpost.com