Retrieval-augmented generation (RAG) typically incorporates external knowledge by appending retrieved documents to the input prompt. While effective for knowledge-intensive tasks, this in-context formulation increases memory use and serving cost as the retrieved context grows. Motivated by this limitation, we propose Doc2Weight, an in-weights RAG framework that injects retrieved information into model parameters rather than the textual context. Doc2Weight converts retrieved documents into parameter-efficient low-rank updates and fuses them in weight space according to their relevance to the query. This enables retrieval augmentation with a query-only prompt, decoupling retrieval depth from prompt length while allowing multiple retrieved passages to influence generation within a single forward pass. We evaluate Doc2Weight on instruction-tuned language models of different scales, including LLaMA-3.2-1B, Qwen2.5-1.5B, and LLaMA-3-8B, across major QA benchmarks. Experiments show that Doc2Weight achieves the best average F1 among standard in-context RAG and prior in-weights RAG baselines across backbones and prompting settings, while matching the throughput of in-context RAG once document signatures are cached. Our code is publicly available.