AnalysisAI ModelsNovember 19, 2026

Meta FAIR's Self-Improving Memory scores 56.0 on continual learning bench

Read original source →marktechpost.com

Self-found memory prompts averaged 56.0 across 3 tasks versus 44.6 for the best human-written ones, and hit 64.3 on Codebase (+20.6 over the best hand-written harness). Built on Kimi-K2.6, a 1T-parameter MoE with 32B active and 256K context, plus a rank-64 LoRA adapter.

1 source

More stories today

Open the live feed