MIT just made every AI company’s billion-dollar bet look embarrassing. They solved AI memory—not by building a bigger brain, but by teaching it how to read. The Breakthrough: On December 31, 2025, three MIT CSAIL researchers published a paper revealing that AI models don’t need massive context windows. Instead of loading entire documents into memory, they store them as external Python variables. The AI knows these exist, searches them using code and regex, pulls only the relevant sections, and spawns sub-AIs to analyze pieces in parallel. No summarization, no information loss. The Problem: Traditional AI models suffer from a hard context window. Overloading it leads to “context rot”—facts blur, mid-document info vanishes, and models forget what they read. Retrieval Augmented Generation (RAG) tried to fix this by chunking documents, but it shredded context and guessed relevancy poorly. The Results: RLMs (Reading-augmented Language Models) solved complex long-context benchmarks where GPT-5 failed 90% of the time. Handled 10 million tokens—100× a model’s native window. Delivered better answers at comparable or cheaper cost. The Implication: For five years, the AI arms race chased bigger windows: GPT‑3 (4K), GPT‑4 (32K), Claude 3 (200K), Gemini 2 (2M). MIT proved the assumption wrong: more context ≠ better performance. The right approach is teaching AI where to look. The Impact: Open source code on GitHub; drop-in replacement for LLM APIs. Enables tasks spanning weeks or months via self-managed context. Ends the context window wars—MIT won by walking away. Sources: Zhang, Kraska, Khattab · MIT CSAIL · arXiv:2512.24601 Paper: arxiv.org/abs/2512.24601 GitHub: github.com/alexzhang13/rlm #mit #gpt #codex #rag #claude #anthropic