Recursive Language Models: Prime Intellect's Paradigm for Agents That Manage Their Own Context Without Summarization

Prime Intellect Blog·high signal

Prime Intellect's blog post argues RLMs—where the model uses a persistent Python REPL and sub-LLM calls to actively manage context rather than summarize it—handle inputs 100x beyond context window size while outperforming long-context scaffolds. Unlike summarization approaches, RLMs preserve information fidelity by delegating to scripts and sub-agents. Training environments are available via prime-rl, making this immediately applicable for long-horizon agent builders.