@0xCodio: A guy in Seoul deleted 68% of his AI's memory file and the answers got better. He didn't guess which lines to cut. He m…

A guy in Seoul deleted 68% of his AI's memory file and the answers got better. He didn't guess which lines to cut. He made OpenAI grade Claude's memory, one line at a time. He published the harness. Forty real prompts from his own history. Each run twice - once with the full file, once with a single line pulled out. 104 lines, 4,160 head-to-head comparisons, $1.40 in judging. The other lab is the part people skip. Ask a model to score its own memory and it defends it. It wrote those lines. It's the author, not the auditor. The result: 71 of the 104 lines never changed a single answer. Not once, across forty prompts. They rode along on every call he made that month. Dead on arrival: "prefers concise answers." "Interested in AI and systems." "Likes clean, readable code." Survived: "ships to production on Thursdays." "Rejected the queue-based version in March, don't re-propose it." Every line that survived was a fact. Every line that died was an adjective. Adjectives feel like memory and narrow nothing. A memory line earns its place only by deleting an answer the model would otherwise have given. Don't let this rot in your bookmarks. The article below is the whole harness - prompt selection, the judge instruction, the diff format that makes the dead lines obvious in one pass. Run it once and you'll cut most of your file tonight.