According to God of Prompt on X, new research comparing 9 memory systems across 1,540 questions finds retrieval methods, not write-time memory strategies, are the dominant driver of AI agent accuracy, with retrieval causing up to 20-point swings while write strategies yield only 3–8 points (as reported by the original X thread). According to the same source, raw conversation chunks with zero LLM preprocessing matched or outperformed fact extraction and summarization pipelines, indicating expensive preprocessing can discard useful context. The thread reports hybrid retrieval combining semantic search, keyword matching, and reranking cut failures roughly in half, and models used relevant context correctly 79% of the time, with retrieval quality correlating strongly with accuracy at r=0.98. For practitioners, this implies prioritizing hybrid retrieval, careful chunking, and reranking over token-heavy write-time compression to boost agent reliability and reduce costs (according to God of Prompt on X).