What it can't remember

Running the model & the world

The idea

Facts live in weights; conversation lives in context. Between them is the thing models still cannot do: learn a new fact for good at inference time.

Karpathy draws exactly this page’s line in his own vocabulary: weights hold a vague long-term recollection; context is exact working memory (ch. 1:32-1:33). And he names the open problem plainly - models today are frozen at deployment, and only the tokens in context ever change at use-time; humans consolidate (his analogy: something like sleep), models don’t, yet (ch. 3:13).

The story, animated

The same fact, two ways - exact in context, vague in weights - and no path ever writes a new one back.

Where it sits

Worth having first: Columns write back · Instruction tuning & RLHF

Episode 5 · What It Can’t Remember

Source-checked against Karpathy, “Deep Dive into LLMs like ChatGPT” (2025).