Wujia / Haibo Wu

Blog 文章

How a Resource-Constrained Open-Source Project Got Discovered 一个没资源的开源项目,是怎么被人看见的

A postmortem on Memex’s open-source cold start: how a project with no budget or partnerships was discovered through real work and useful discussions.

Memex 开源冷启动复盘:没有预算、没有合作,如何靠真实作品和有效讨论被看见。

English translation provided here; the Zhihu link is the original source. 中文为知乎原文,英文为译文。

Recently, Memex was recommended by a few overseas AI bloggers.

One Instagram post got 16,000 likes. The project picked up some GitHub stars as well, though it was still far from “going viral.”

The funny part is that at first we had no idea what had happened.

GitHub analytics only showed a few vague referrers. We searched X, TikTok, and YouTube ourselves but still could not find the real source. In the end I asked ChatGPT to search whether anyone overseas had mentioned Memex recently, and that was how I found those recommendation posts.

For an AI project, discovering the source of growth through ChatGPT is, in a way, very GEO.

https://www.instagram.com/reels/DZFafmFNwMp/

一个没资源的开源项目,是怎么被人看见的

Memex is an open-source, local-first AI journal. You can drop in fragments of life — text, images, voice, screenshots — and let AI organize them into your own memory.

But this piece is not mainly about product features. I want to record how a small open-source project with no resources, no paid promotion, and no partnerships slowly became visible to other people.

The earliest starting point was not the product itself, but our research into Claude Code.

We realized early that agents do not necessarily need complex databases, indexes, or RAG from day one. In many cases, if information is organized as markdown and paired with basic bash commands such as find, grep, and file reads and writes, an agent can already do very powerful things.

That judgment later became the underlying intuition behind Memex’s early architecture.

We were not building a traditional diary app. We wanted to build a “file system for personal memory”: records can be fragmented, organization can be handed to agents, and structure can grow gradually over time.

Of course, when we first released it, almost nothing happened. For example, I wrote a product-thinking piece titled “We Built Something About ‘Recording Yourself,’ Then Decided to Open Source It,” and it barely got any traffic.

So we wrote bits of thinking on Xiaohongshu and Zhihu, kept improving the README, built the website, shipped to the App Store and Google Play, and kept using and improving the product ourselves. That stage was slow, and it was easy to wonder whether we were the only ones who found the idea interesting.

The first wave of cold-start growth actually came from a comment.

A while ago, Karpathy published a gist called LLM Wiki. The rough idea was to use LLMs to maintain a markdown-based personal knowledge base. Instead of the traditional RAG pattern of “retrieve only when a question is asked,” the model would continuously organize, update, and connect markdown files, allowing the knowledge base to grow by itself over time. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

It resonated strongly with us, because it was very close to many of the ideas behind Memex.

So when we saw Karpathy’s gist, we left a serious reply in the comments and mentioned Memex.

That became our first wave of cold-start traffic.

The experience left a strong impression on me: when you have no resources, the best promotion may not be “promotion” at all, but showing up in the right conversation.

You have to have really thought about the problem and really built something, so that when others are discussing it, you can meaningfully join the conversation.

This is very close to built in public.

It is not about shouting every day that “we shipped another update.” It is about continuously exposing your judgments, iterations, and confusions. Early on, there may be almost no feedback, but those traces accumulate. When a discussion happens to intersect with your work, people can follow those traces back to you.

Later, Memex slowly crossed 100 stars.

一个没资源的开源项目,是怎么被人看见的

That number is not large, but for a small open-source project it felt like a faint signal: at least the thing was not completely ignored. After that point, a few smaller bloggers found us, then somewhat larger accounts reposted the project, and eventually there was that Instagram post with 16,000 likes.

So the path was actually very plain:

First, we clarified some of our own thinking. Then we built the product. Then we spoke seriously in relevant discussions. Then the first group of people came in. Then we kept iterating. Then more people happened to discover it.

There was no magical method.

This was a bootstrapped project. We had no budget, no partnerships, and no overseas channels. A lot of the work was simply slow accumulation: improving the README, adding to the website, writing on Xiaohongshu and Zhihu, shipping to the App Store and Google Play, responding to user feedback. Gradually it became a product serving users around the world.

一个没资源的开源项目,是怎么被人看见的

This experience made me more certain of one thing: in this stage of AI agents, individual developers do have opportunities.

A lot of things are being re-understood. You may not have resources, but if you understand a new paradigm early enough, keep turning ideas into working products, and iterate in public, you may be discovered by others who are thinking about the same thing.

Of course, the other side is painful.

Our understanding of agents has changed quickly over the past few months. Once your understanding changes, you start seeing many awkward parts in the old product architecture. Memex is now preparing for a fairly large refactor.

That is also what I find interesting about built in public: it is not just making an outcome public, but making public how an idea changes, and how a product is pushed forward by your own new understanding.

Memex is still very early, and many parts are not good enough. But if you are interested in AI agents, local-first software, personal memory, or companionship-like products, you are welcome to visit GitHub or join the work.

Project: memex-lab/memex

Website: Memex — Open-Source AI Journal That Organizes Your Life

最近 Memex 被几个海外 AI 博主推荐了一下。

其中一条 Instagram 帖子有 1.6 万 likes。项目本身也跟着涨了一波 stars,但还远远谈不上火。

比较好笑的是,我们一开始完全不知道发生了什么。

GitHub 后台只能看到一些很模糊的来源。我们自己去 X、TikTok、youtube 上找了一圈,也没找到真正的原因。最后我是问 ChatGPT,让它帮我搜最近有没有人在海外提到 Memex,才发现原来有几条帖子在推荐我们。

一个 AI 项目的增长来源,最后靠 ChatGPT 查出来,多少有点GEO。

https://www.instagram.com/reels/DZFafmFNwMp/

一个没资源的开源项目,是怎么被人看见的

Memex 是一个 open-source、local-first 的 AI journal。你可以把文字、图片、语音、截图这些生活碎片丢进去,让 AI 帮你整理成自己的记忆。

但这篇不太想讲产品功能。我更想记录一下:一个没资源、没投放、没合作的小开源项目,是怎么慢慢被别人看见的。

最早的起点,其实不是产品,而是我们对 Claude Code 的研究。

我们比较早意识到一件事:Agent 不一定一上来就需要复杂的数据库、索引、RAG。很多时候,只要把信息组织成 markdown,再配合基础的bash命令,如find、grep、读写文件,Agent 就已经能做非常强的事情。

这个判断后来变成了 Memex 早期架构的底层直觉。

我们做的不是一个传统日记 App,而是想做一个“个人记忆的文件系统”:记录可以很碎,组织可以交给 Agent,结构可以慢慢长出来。

当然,刚发出来的时候,基本没什么反应。比如我写了一篇产品理念:我们做了一个关于"记录自己"的东西,然后决定把它开源 一点流量都没有。

我们就在小红书、知乎写一点想法,改 README,做官网,上 App Store 和 Google Play,一边自己用一边改。这个阶段很慢,也很容易怀疑:是不是只有我们自己觉得这件事有意思?

第一波冷启动,反而来自一次评论。

前段时间 Karpathy 发了一个 gist,叫 LLM Wiki。大概意思是,用 LLM 去维护一个基于 markdown 的个人知识库。不是传统 RAG 那种“每次问的时候再去检索”,而是让模型持续整理、更新、链接这些 markdown 文件,让知识库自己慢慢长出来。https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

这个东西我们看了之后很有感觉,因为它和我们做 Memex 时的很多想法非常接近。

所以当时看到 Karpathy 那个 gist,我们就在评论区认真回了一条,也放了 Memex。

这就是第一波冷启动。

这件事给我的感受很强:
当你没有资源的时候,最好的推广可能不是“推广”,而是出现在正确的讨论里。

你得真的想过这个问题,真的做过一些东西,然后在别人讨论到这里的时候,你能接得上。

这很像 built in public。

不是每天喊“我又更新了什么”,而是持续把自己的判断、迭代、困惑放出来。早期不会有什么反馈,但这些东西会变成痕迹。等某个讨论刚好和你有关的时候,别人能顺着这些痕迹找到你。

后面 Memex 就慢慢过了 100 stars。

一个没资源的开源项目,是怎么被人看见的

这个数字不大,但对小开源项目来说,它像一个很微弱的信号:至少这个东西不是完全没人关心。过了这个点之后,开始有一些小博主看到我们,再后来有稍微大一点的账号转发,最后才有了那条 1.6 万赞的 Instagram 帖子。

所以这条路径其实很朴素:

先是我们自己想清楚一点东西。
然后把产品做出来。
然后在相关讨论里认真说话。
然后有第一波人进来。
然后持续迭代。
然后被更多人偶然看见。

没有什么很神奇的方法。

这个是业务项目,我们没有预算,没有合作,也没有海外渠道。很多事情就是慢慢做:README 改一改,官网补一补,小红书知乎写一写,App Store 和 Google Play 上一上,用户反馈修一修,慢慢的它又变成了一个服务全球各地用户的产品了。

一个没资源的开源项目,是怎么被人看见的

这次经历让我比较确定的一点是:AI Agent 这个阶段,对个人开发者是有机会的。

因为很多东西大家都还在重新理解。你没有资源,但如果你对一个新范式的理解足够早,持续把想法做成东西,并且公开迭代,就有可能被同样在思考这件事的人看到。

当然,另一面是痛苦的。

我们这几个月对 Agent 的理解变化很快。理解变了以后,就会发现老的产品架构有很多地方不对劲。现在 Memex 也在准备一轮比较大的重构。

这也是我觉得 built in public 有意思的地方:它不只是公开一个结果,而是公开一个想法怎么变化,一个产品怎么被自己的新认知反过来推动。

Memex 还很早期,也有很多做得不好的地方。
但如果你也对 AI Agent、local-first、个人记忆、陪伴这类方向感兴趣,欢迎来 GitHub 看看,或者一起参与进来。

项目:memex-lab/memex

官网:Memex — Open-Source AI Journal That Organizes Your Life