Open-Source Memory Engine NylonME Doubles LoCoMo Recall Score from 47% to 84%

The NylonME memory engine improved its LoCoMo evidence-recall score from 47.1% to 84.6% over two weeks of iterative development, as documented by its developers. The initial baseline relied solely on lexical matching, which failed to retrieve facts phrased differently from the query. Key improvements included integrating bge-m3 vector embeddings for semantic retrieval, a dual-layer write architecture that preserves both raw conversation text and LLM-distilled facts, and a final vector reranking step. An ablation test revealed that using only LLM-distilled abstract facts actually dropped the score to 67.3%, underscoring the importance of retaining verbatim conversation data. The developers published the full record of both successful changes and failed experiments, citing a lack of transparent benchmarking in the memory-systems field.