deepseek-ai/Engram

Why this repo matters

• •Introduces conditional memory via scalable lookup as a complementary sparsity axis to Mixture-of-Experts architectures.
• •Demonstrates consistent improvements over MoE baselines under strict iso-parameter and iso-FLOPs constraints across multiple domains.
• •Enables efficient host memory offloading for massive embedding tables using deterministic addressing with minimal inference overhead.