Large windows can accept substantial documents or much of a project. Context capacity does not guarantee that a model uses every part equally well.
About this eventThe workaround the industry lived on while context was short. Now a complement rather than a crutch.
RAG separates knowledge from model parameters: before answering, a system retrieves relevant documents and adds them to the input. This makes facts and citations easier to update, but quality depends on retrieval, document chunking and the model's ability to distinguish evidence from its own guess.
Source: arXiv · RAGAbout this eventA million tokens in production — a whole book or an hour of video fits inside.
About this eventCoding systems read the whole project.
About this eventOne-twentieth the compute per token, 9× prefill and 15× decode on million-token context.
About this eventAAIF reviewed file memory, retrieval, profiles and checkpoints. This dates the overview, not memory's invention; these systems do not necessarily update model weights.
About this eventLong context fits, but attention to its middle sags: the model recalls the start and the end better than what lies between.
About this eventCross-session memory is already used. Freshness, forgetting, access control and checking retrieved information remain open challenges.
Primary material behind this dossier: papers, lab publications and official reports.