DeepSeek releases a new paper by Liang Wenfeng on the subject of memory module Engram, with his signature.

DeepSeek released a new paper "Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models" on the evening of the 12th. The paper was completed jointly by Peking University and DeepSeek, with co-authorship by Wenfeng Liang. The paper proposes conditional memory, which significantly improves the performance of models in tasks such as knowledge retrieval, reasoning, code, and mathematics under the same parameter and computational conditions by introducing a scalable lookup memory structure. At the same time, DeepSeek has open-sourced the related memory module Engram.