Stanford researchers open-source Memorilla, a memory module for LLMsMachine translation

Memorilla turns a document collection into 16 memory tokens for a frozen LLM. In the research team's self-reported FactKG results with Qwen3-8B, it scored 84.33% versus 80.43% with the full documents in context, using 108 times fewer document tokens; it trailed a comparison method on PubMedQA. Code, training recipes, and checkpoints are public.