Shashi Jagtap explains the problem with loading an entire large repository into a model context. Cost rises, relevant evidence becomes harder to find and summaries may discard precisely the detail needed for a correct code change.
The recursive language model pattern stores the repository as an external variable and lets the model write code that searches, slices and computes over it. The model can ask focused subquestions, inspect results and repeat the process before producing a final answer.
A live RLM Code demonstration shows the value of bounded queries, replayable trajectories and cost tracking. Jagtap presents the method as a research harness for comparing strategies, not a guarantee that recursion alone will solve every coding task.
Watch on YouTube
