The researchers proposed a memory for the LLM memorandum: it reads the entire document database and aggregates the desired request in 16 vectors. The documents are encoded once by Qwen3-Embedding-4B, and the 194 million dm trainer feeds these 16 vectors directly into Qwen3-8B input symbols. According to the authors' measurements, the method leads 5 out of 8 measures among the versions of RAG, Mem0, MemoRAG and full context. In FactKG, the accuracy is 84.33% compared to 80.43% in full context while reducing the document part of the input by 108 times. But on PubMedQA Memorilla 59% vs. 67-69% for best practices, and on NarrativeQA (ROUGE) and PersonalizationV4 before MmoRAG. The module was also trained on training parts of the same 8 measures, and small details during compression, according to the authors, can be lost. Also, you need access to embedding the input of the model, so don't use the normal text API method. Channel @neuro.

The researchers proposed a memory for the LLM memorandum: it reads the entire document database and aggregates the desired request in 16 vectors. The documents are encoded once by Qwen3-Embedding-4B, and the 194 million dm trainer feeds these 16 vectors directly into Qwen3-8B input symbols. According to the authors' measurements, the method leads 5 out of 8 measures among the versions of RAG, Mem0, MemoRAG and full context. In FactKG, the accuracy is 84.33% compared to 80.43% in full context while reducing the document part of the input by 108 times. But on PubMedQA Memorilla 59% vs. 67-69% for best practices, and on NarrativeQA (ROUGE) and PersonalizationV4 before MmoRAG. The module was also trained on training parts of the same 8 measures, and small details during compression, according to the authors, can be lost. Also, you need access to embedding the input of the model, so don't use the normal text API method. Channel @neuro.