MIT CSAIL researchers Alex Zhang, Tim Kraska and Mohammad Aliannejadi Khattab dropped a paper on December 31, 2025 that rewrites how AI handles long documents. Instead of cramming text into a model’s context window and watching it decay through “context rot,” they keep the entire document as a Python variable outside the AI’s memory. The AI then writes code to grep, slice and navigate exactly where it needs to look. When it finds relevant sections, it spins up focused sub-instances to read and analyze those parts in parallel before the main model merges their findings. No chunking, no blind retrieval guesses, no lost links between distant paragraphs. On benchmarks that push models beyond 75,000 tokens, standard LLMs collapse—GPT-5 solves fewer than 10 percent of the problems. The new Retrieval-Linked Models (RLMs) not only handle inputs up to ten million tokens, they deliver double-digit percentage gains over Retrieval-Augmented Generation methods at equal or lower cost per query. RLM’s code is open source on GitHub; you swap out your existing API calls and get the same interface, but suddenly your app works on archives it used to choke on. With this shift, the era of betting on bigger context windows—GPT-3’s 4,000 tokens to Gemini 2’s two million—looks like a dead end. RLMs teach AI to know where to look instead of how much to remember.