DeepSeek Engram: Conditional Memory via Scalable Lookup — New Sparsity Axis for LLMs — 4,111 Stars
GitHub·low signal
DeepSeek's Engram research introduces conditional memory as a complementary sparsity axis alongside MoE, identifying a U-shaped scaling law for optimal capacity allocation. The Engram-27B model shows gains across knowledge (MMLU +3.4), reasoning (BBH +5.0), and code/math (HumanEval +3.0) domains under strict iso-parameter constraints. Deterministic addressing enables runtime prefetching with negligible overhead. Paper from January 2026, code continues to trend.