Stanford Open-Sources Memorilla Memory Module for Frozen LLMs
Stanford researchers have released Memorilla, an Apache 2.0 open-source latent memory system for frozen LLMs. A 194M-parameter trainable module reads a document collection and supplies a frozen Qwen3-8B decoder with 16 memory tokens while the Qwen3-Embedding-4B encoder and decoder remain frozen. On FactKG, Memorilla scored 84.33% versus 80.43% for full-context prompting with 108x fewer document tokens, and it reported the highest average across eight benchmarks at 49.68. The authors note weaker results on PubMedQA and fine-grained factual details, and the released checkpoints are paired with specific Qwen decoders.
Why it matters: It gives readers a concrete, open-source memory architecture with published benchmarks, deployment details, and documented limitations for extending frozen LLMs.