DeepSeek Engram sparks V4 speculation

This is early-stage community inference, not an official roadmap. The interesting part is that the speculation is anchored in a real DeepSeek research repo, so the rumor has technical footing even if the release timing does not.

• –The Engram repo describes conditional memory via scalable lookup, with static N-gram memory fused into the backbone for O(1) retrieval.
• –DeepSeek frames Engram as a sparsity axis alongside MoE, which makes it relevant to long-context and efficiency tradeoffs.
• –Tno official confirmation here that future public models will ship "updatable engrams"; treat that as forum speculation.
• –If DeepSeek productizes this idea, the bigger impact is likely lower inference cost and better retrieval behavior, not just a bigger benchmark headline.