AI models that read and write text, known as large language models or LLMs, have a costly habit. They use the same expensive processing to recall a basic fact as they do for hard reasoning. Asked for the capital of France, a model still runs its full reasoning machinery, though the answer is a simple lookup.
Much real-world AI use consists of routine questions and repeated patterns like that one, asked in huge numbers. Each one costs more computing and time than it needs to, and the waste grows as use rises and models get bigger.
A DeepSeek paper traces the habit to a missing part, a fast and dedicated way to recall fixed facts. People have one, which is why nobody works out 2+2 from scratch, and the paper argues models should work the same way. Its fix, a component called Engram, gives models that separate path for memory. A model can then fetch stored patterns cheaply and keep its costly computing for real reasoning.
The split extends to the hardware. Certain stages of the model connect to large tables of stored patterns, which sit on cheaper hardware such as a server’s ordinary memory. The most expensive AI chips can then focus on the core calculation. The aim is recall that is fast, predictable and cheap.
To test it, DeepSeek trained large Engram models and compared them with strong standard models that use the same amount of computing. As expected, Engram did better on tests of factual knowledge. More surprisingly, it made larger gains on tests of math, coding and multi-step logic.
The paper’s analysis suggests a reason. Once the model stops “relearning” simple patterns in its early stages, more of its processing is free for harder logic. The design also did well on tests built around long documents and finding information in them. Such tests resemble work on contracts, policies and long emails, though DeepSeek did not test those tasks directly.
Design choices like this set the cost of each AI answer and how many answers a system can produce. For any company using AI in earnest, the cost question is how much of its AI use pays “reasoning prices” for work that is really recall.
