Memorilla preprint reports improved retrieval using 16 memory tokens

Researchers report that Memorilla, a memory system for language models with fixed parameters, outperformed baselines including Mem0 and MemoRAG on five of eight benchmarks using 16 memory tokens. The arXiv preprint covers personalization, long-document questions and knowledge-intensive questions. Memorilla encodes documents once into reusable numerical representations, then learns to place query-specific information into the model's input. On FactKG, it reached 84.33% accuracy versus 80.43% with full-context inference, using 108x fewer tokens to represent source documents. In TextWorld, average agent success rose from 18.6% with full-history input to 24.2%; these results concern the evaluated benchmarks and environment.