Part of a blog series on memory consumption and limitations in LLMs with large context windows. Here, we explore tokens, embeddings, and memory.| Revelry