Chunk the raw conversation into 256-token windows, retrieve the 128 nearest by embedding, rerank down to 15, hand them to the reader. No fact extraction, no session summaries, no graph. Against twelve purpose-built memory systems across four long-conversation benchmarks, that ranked first in 6 of 7 benchmark-and-model settings, by 0.13 to 5.53 points. The reason is not cleverness: across systems, how often the retrieved evidence was sufficient at all swings 26 points, while accuracy given sufficient evidence swings under 10.