初中生物课上，老师说，语言是人类区别于其他动物的重要特征之一。历史课上，我们学到蔡伦改进造纸术，以及课本里颇为自豪的中国四大发明。当时，我对这些知识的关注主要还停留在考试范围内。
从纸张到印刷机、打字机，再到计算机，知识的记录、复制和保存变得越来越容易。我出生时，互联网正在普及。一个人写下的想法，可以被世界另一端的陌生人读到，Google 又让这些散落各处的内容有了统一的搜索入口。
大二时，我开始接触 ChatGPT。最初的用途很实际，临近上课，让它帮忙翻译、总结还没读完的材料。它有没有帮助我学习，不好说，但至少在当时是救命稻草。
此后，模型处理上下文的能力不断提升。通过 RAG、MCP 等技术和协议，我们也能在不重新训练模型的情况下，为它提供所需的信息。模型可能熟悉软件工程的各种原则，却不知道我们团队为什么做出某个具体的设计决策。通用知识无法替代这些背景，前提仍然是有人把它们记录下来，并让模型能够访问。
这让文档有了更大的价值。关闭 ServiceNow 工单时，我们一直都需要填写解决方案，供后来遇到类似问题的人参考。过去，复用这些经验往往需要猜关键词、翻历史工单。现在，AI 可以检索相似故障、比较处理方式，再根据已有记录提出建议。对于生产支持而言，每一次排查留下的记录，都可能减少下一次的重复工作。
个人记录也是如此。尝试过的护肤品、正在推进的目标、做某个决定时的考虑，都能为 AI 提供具体的背景，让建议更贴合自己的情况。我们不必每次重新交代一遍，它也不必只凭通用知识回答。那些自己已经记不清的细节，还可以从记录中找回来。
记忆开始变得可以查询。
这让我重新想到初中生物课上的那句话。当时，语言意味着人类能够彼此交流，让知识跨越世代。我没有想到，这些积累下来的文字，有一天会帮助我们构建出能够向我们学习、与我们交流的系统。我们一直通过语言把知识传给后来的人，未来的接收者，或许也会包括一种由我们自己创造的智能。
In middle school, my biology teacher told us that language was one of the things that made humans different from other creatures. In history class, we learned about Cai Lun and papermaking, proudly listed as one of China’s four great inventions. At the time, these were mostly things I needed to remember for exams.
Then came printing presses, typewriters, and eventually computers. Each made some part of creating, copying, or storing information easier. By the time I was born, the internet was becoming part of everyday life. People could publish something online and reach readers they would never meet. Google made much of that information searchable.
During my sophomore year of college, I started hearing about ChatGPT. My first use case was fairly practical — summarizing readings I had left until shortly before class. Whether this improved my education is debatable, but it helped with the immediate problem.
Since then, models have gotten better at using context, and we’ve built ways to supply it through approaches like RAG and protocols like MCP, without retraining the model. A model might know a lot about software engineering, but understanding why our team made a particular choice still depends on someone recording that decision and making it available.
Documentation has always been useful at work. We write resolution details when closing ServiceNow tickets so someone facing the same issue can reuse the fix. AI makes that reuse much easier: it can retrieve incidents with similar symptoms, compare previous resolutions, and suggest a fix. For production support, a history of solved problems becomes a resource for solving the next one.
The same applies to personal life. Notes about products I’ve tried, goals I’m working toward, or decisions I’ve made give AI context for more relevant suggestions without requiring me to explain everything again. They also preserve details I would otherwise forget. As more of that experience is recorded and made searchable, memory becomes queryable.
This brings me back to my middle school biology class, where I learned that language was one of the things that made us human. I didn’t imagine that what we wrote down could one day help us build something that could learn from us and respond. It raises a possibility I hadn’t considered. Language might allow us to pass knowledge on to a kind of intelligence we created ourselves.
