A 16-token memory approach reportedly outperforms full-context inference on FactKG

A post summarizing a paper on latent semantic memory says its approach encodes documents once into reusable embeddings, then writes query-relevant information into 16 tokens for a frozen LLM. The post reports a FactKG score of 84.33%, versus 80.43% with full-context inference, while using 108 times fewer tokens for the source documents. It says the approach beat Mem0, MemoRAG and other baselines on five of eight benchmarks.