LambdaDB published a memory pipeline that stores conversation passages and retrieves relevant chunks through embeddings, hybrid search, and reranking. Without building a knowledge graph or using a generative model to summarize and maintain memories, it scored 96.0% on 500 LongMemEval-S questions versus 95.2% with full history. Mean answer input fell from 105,183 to 4,886 tokens, a 95.35% reduction. The implementation is available as open-source SillyMemory, but these are vendor-run tests of English configurations. Its official repository describes an experimental release and advises starting with a test chat and separately checking model and browser compatibility.