🧠Memorilla Packs Whole Document Collections Into Just 16 Memory Tokens

Your frozen LLM can finally remember without retraining

A new paper called Memorilla lets frozen LLMs remember whole document collections using just 16 memory tokens. No fine-tuning, no retraining, and the base model stays untouched. That's the part that matters. If you're shipping RAG or agents on a frozen model, your context budget is the bottleneck. Memorilla encodes documents into reusable embeddings, then learns to retrieve query-specific information and inject it into the model's input embedding space. Think of it as a smarter memory layer instead of stuffing raw chunks into the prompt. The agent side is interesting too: in a TextWorld test, it wandered between already-explored rooms less and improved its average success rate. The numbers: 84.33% accuracy on the FactKG benchmark, using 108x fewer tokens to represent the source documents. It beat baselines on five of eight benchmarks with those same 16 tokens. The paper was submitted October 15, 2026, and revised November 3, 2026, with three versions out. Watch whether the gains hold when document counts grow through a long agent trajectory. The paper frames this as a step toward continual learning, not a finished solution.