Prime Intellect’s research post of 1 January 2026, Recursive Language Models: the paradigm of 2026, calls the RLM the simplest and most flexible form of context folding: instead of ingesting its input, the model uses a persistent Python REPL to inspect and transform the data and to call sub-models from inside the REPL, so PDFs, datasets or videos never have to be loaded into the context at all (blog). The post credits the idea to Alex Zhang’s October 2025 blog post and the MIT paper published the day before, and says the RLM is now a major focus of the lab’s research; the corpus records the paper’s own result, contexts two orders of magnitude beyond the window, two days later. In the recursive-self-improvement trajectory it is the first entry in which a model administers its own working memory, the scaffold-side twin of Stanford’s end-to-end test-time training and a precursor of agents that meta-learn their own memory architecture.
Development
A model takes over its own context via a Python REPL
The human leaves the context window: a lab names as its paradigm for 2026 a scheme in which the model, not the operator, decides what enters its own working memory.
Claim
Prime Intellect unveiled a Recursive Language Model that manages its own context through a persistent Python REPL, inspecting and transforming data end-to-end without human oversight.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to January 2, 2026 (theinnermostloop.substack.com)
- Recursive Language Models: the paradigm of 2026 (www.primeintellect.ai)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.