本文由 AI 分析生成
建立時間： 2026-01-02 來源： https://x.com/op7418/status/2006937416125526499
Summary
The author highlights how novelist Ma Boyong and Andrej Karpathy both keep diaries by recording only facts (no opinions, no polishing) in a single append-only file, searched via Command+F. This low-friction, high-density, plain-text style — capturable by voice — turns out to be excellent AI-native corpus: small enough to fit a model’s context window, semantically dense (names, books, events, commands, code), and not locked to any product. The author plans to adopt it in 2026 to build a personal ChatGPT-like memory.
作者點出小說家馬伯庸與 Andrej Karpathy 都用「只記事實、不寫感悟、不潤色」的方式寫日記，全部丟進單一 append-only 檔案，用 Command+F 檢索。這種低摩擦、高密度、純文字（可語音輸入）的風格，意外地非常適合 AI：體積小到能塞進模型上下文、語義密集（人名、書名、事件、指令、程式碼），且不被任何產品綁定。作者計畫在 2026 年實踐，用它建立個人化的類 ChatGPT 記憶。
Key Points
- Record only facts (“流水账”) — no opinions, no perfomance, no polishing — to slash input friction.
- Append-only single file; no tags, no categories, no hierarchy; search with Command+F.
- A year of such notes is only tens of thousands of words — small, retrievable, context-friendly.
- Ma Boyong uses it against middle-age memory decline; Karpathy treats it as releasing brain RAM.
- Plain text avoids product lock-in and is an ideal AI-native corpus.
Insights
Removing opinion and polish does double duty: it kills the friction of self-presentation and produces cleaner structured data for an LLM to extract. The single-file, append-only constraint is the same “brain as CPU, not RAM” idea seen in task-management circles — offload state so the mind can focus. The deeper bet is that low-friction plain text plus a long-context model gives you a private, portable memory layer without committing to a proprietary memory feature.
Connections
Raw Excerpt
这种低摩擦、高信息密度，且纯文本的记录方式，无意中非常适合现在的 AI。