Teaching frozen LLMs new knowledge without touching their weights

LLMs go stale the moment training ends, yet most update strategies require either full retraining, access to model weights, or retrieval systems that slow down as document collections grow. MeMo sidesteps all three by training a small, dedicated memory model on the new corpus and attaching it to an unchanged LLM at inference time. The memory model learns cross-document relationships rather than treating each passage independently, making it resilient to retrieval noise. On BrowseComp-Plus, NarrativeQA, and MuSiQue, MeMo matches or outperforms existing retrieval-augmented and continual-learning baselines.