alexzhang13/rlm: Recursive Language Models Library — Models That Call Themselves for Near-Infinite Context, 3.5K Stars

GitHub·medium signal

RLM introduces a novel inference paradigm where language models offload context into a REPL environment and recursively launch sub-LM calls to decompose complex tasks. This enables handling near-infinite context lengths without traditional context window limits. Supports OpenAI, Anthropic, and OpenRouter with sandbox backends including local, Docker, Modal, E2B, and Daytona. Includes a web-based trajectory visualizer for debugging execution paths. At 3.5K stars, it's an early but genuinely novel approach to the context length problem.