Preprint proposes lazy memory retrieval for long LLM chats

A preprint describes Pull, a session router intended to reduce the context supplied to stateful LLM conversations. It maintains a local metadata directory and retrieves only turns needed for a query, while leaving collapsed turns available for later expansion. On the reported LoCoEval tests, Pull reduced per-query context tokens by 75.1% on single-hop tasks and 72.0% on multi-hop tasks with no reported quality loss. The results are from controlled benchmarks, not a production deployment.