DeepSeek's Engram adds conditional memory to LLMs

This feels less like a clever tweak and more like an attempt to give sparse LLMs a real memory plane: let the model store reusable facts cheaply, and spend neural depth on actual reasoning. If those results generalize, Engram is the kind of architectural idea that could outlive a single model family.

• –It cleanly splits conditional compute from conditional memory, which is a more interesting scaling axis than simply adding more experts.
• –The strongest signal is breadth: the paper claims wins on knowledge, reasoning, code, math, and long-context tasks, not just factual recall.
• –Deterministic addressing and host-memory prefetching make the efficiency story compelling for serving, especially when long-context throughput matters.
• –The big risk is operational: memory collisions, table growth, and retrieval quality will decide whether Engram stays a research win or becomes a production primitive.