Everything You Need to Know About Recursive Language Models

Recursive Language Models (RLMs) address the 'context rot' problem where LLMs degrade in quality when given very long inputs. Instead of feeding the entire prompt into a single forward pass, RLMs treat the prompt as an external variable and let the model interact with it through a persistent REPL environment. The model receives only metadata about the prompt, then issues code-based commands to inspect, decompose, and recursively sub-query specific slices of the input. Intermediate results are stored in the environment rather than the model's context window, allowing processing of inputs far exceeding any single context limit. RLMs differ from RAG (which pre-selects relevant chunks) and agent systems (which inject full history into context) by keeping the prompt external throughout and using true programmatic recursion. Tradeoffs include higher orchestration complexity, reliance on the model's code-writing ability, and potentially higher cost variance compared to a single large-context call.