Ralph Loop Explained: Why Clean Context Windows Beat Long AI Sessions

The Ralph loop, named by developer Geoffrey Huntley in July 2025, is a coding agent pattern where each iteration starts with a fresh context window while durable state is stored on the filesystem rather than in conversation history. Huntley designed it to counter the well-documented problem of model performance degrading as context length grows, noting that usable quality drops off well before a model's advertised context limit. The technique is more nuanced than its viral one-line Bash summary suggests — the real principle is keeping the working context small and offloading memory to external files like git history and spec documents. OpenAI has since productized a similar approach in its Codex CLI via a looping goal command, signaling that the pattern has practical merit beyond social media buzz. However, the approach carries trade-offs, including higher API cache costs from repeatedly re-reading filesystem state, which proponents often leave unmentioned.