Meta FAIR memory method scores 56.0 on continual learning bench
Original title: Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench
Optimized prompts beat the best hand-written harness by 11.4 points across three Kimi-K2.6 tasks.
Meta FAIR introduced Self-Improving Memory, a two-stage method that searches an agent’s file-memory prompts and then distills the behavior into weights. Using Kimi-K2.6, self-found prompts averaged 56.0 on Database, Codebase, and Spectrum, versus 44.6 for the best human-written prompts. After LoRA distillation, the model running generic prompts averaged 43.1, above 37.6 for supervised fine-tuning on generic trajectories. Co-authors include researchers from the University of Washington and Johns Hopkins University.
Unconfirmed: Comparisons with ALMA, MemEvolve, and Harness-Zero use different benchmarks, so headline scores are not directly comparable.
- The base model is Kimi-K2.6, a 1T-parameter mixture of experts with 32B active parameters.
- Codebase reached 64.3 with optimized prompts, 20.6 above the best hand-written harness.
- Thirty-seven of 44 generated revisions failed to beat the previous best.
- Distillation uses a rank-64 LoRA adapter on attention projections.
- Training hardware was not disclosed.
- AI models
- Kimi-K2.6
From the original article
8 min read at marktechpost.comMeta FAIR’s Self-Improving Memory lets Kimi-K2. 6 write its own LOAD and OFFLOAD memory prompts, then distills them into its weights. Optimized prompts score 56.
By Michal Sutter · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.