Show HN：一项赋予编码代理长期记忆的实验
1 分•作者: yacc2•5 个月前
最近，我想在实现人们所说的AGI（通用人工智能）的道路上更进一步。<p>很多朋友告诉我这不可能。归根结底，这些模型只是梯度下降和概率机器——基本上是一个非常复杂的马尔可夫链。当然，我是在夸张，哈哈，但你懂我的意思。<p>尽管如此，我还是想看看，围绕模型构建更好的基础设施是否能推动智能体更接近于成为可靠的独立编码员，而不是患有严重失忆症的助手。<p>在阅读了大量关于这个主题的论文后，我开始尝试为智能体构建持久记忆和引导式学习。<p>这个想法很简单：<p>智能体执行的每个动作都会被存储。当未来的智能体处理同一个代码库时，它们会收到从过去的成功和失败中提炼出来的“技巧和窍门”。<p>这些记忆会被嵌入，这样监督层就可以语义搜索相关上下文，并将其注入到提示中，而不会撑爆上下文窗口。<p>我还在尝试让智能体彼此沟通，以便多个任务可以并行运行——即使它们在代码上重叠——并自行解决冲突。<p>为了测试这一点，我建立了一个小网站。<p>目前它还非常简陋，但你可以连接：<p>一个GitHub代码库<p>一个Anthropic API密钥<p>Linear问题<p>然后，你在“待办事项”栏中创建标记为“gradient-agent”的问题，智能体就会开始处理它们。<p>你对代码库的迭代越多，“学习”就会积累得越多。理论上，这应该能使未来的PR（Pull Request，拉取请求）更加稳定。<p>长远目标是建立一个可以自主规划、设计、编码、测试和审查整个PR的系统，而无需持续提示。<p>现在是不幸的部分。<p>我是一个贫困的大学生，这种基础设施并不便宜，哈哈。我添加了每月免费额度，以便人们可以测试它，但一旦这些额度用完，你就要么等待重置，要么按需付费。（LLM API费用仍然由你承担。）<p>另外：这还是一个测试版，所以请谨慎处理极其复杂的任务。<p>我正在考虑托管类似qwen3.5:122b的东西，这样人们就可以在没有自己的API密钥的情况下进行测试。额度会更快用完，但总体上可能更便宜。好奇人们是否需要这个。<p>网站：
<a href="https://usegradient.dev" rel="nofollow">https://usegradient.dev</a><p>最初我计划将测试限制在约20人，但老实说，我很好奇它在更多用户下的表现。<p>拜托，请不要DDOS我。<p>欢迎任何反馈。
查看原文
Recently I wanted to take a couple steps closer to what people call AGI.<p>A lot of friends tell me it’s impossible. At the end of the day these models are just gradient descent and probability machines — basically a very sophisticated Markov chain. Obviously I’m exaggerating lol, but you get the idea.<p>Still, I wanted to see if better infrastructure around the model could push agents a little closer to being reliable individual coders instead of assistants with severe amnesia.<p>After reading a bunch of papers on the topic, I started experimenting with persistent memory and guided learning for agents.<p>The idea is simple:<p>Every action an agent takes gets stored. When future agents work on the same repo, they receive condensed “tips and tricks” derived from past successes and failures.<p>These memories are embedded so a supervisor layer can semantically search for relevant context and inject it into the prompt without blowing up the context window.<p>I'm also experimenting with letting agents communicate with each other so multiple tasks can run in parallel — even if they overlap in code — and resolve conflicts themselves.<p>To test this, I built a small site.<p>Right now it's very bare bones, but it lets you connect:<p>a GitHub repo<p>an Anthropic API key<p>Linear issues<p>Then you create issues labeled gradient-agent in the To-Do column and the agents start working on them.<p>The more you iterate on a repo, the more “learning” accumulates around it. In theory this should make future PRs more stable.<p>The long-term goal is a system that can plan, design, code, test, and review an entire PR on its own without constant prompting.<p>Now the unfortunate part.<p>I’m a poor university student and this infrastructure is not exactly cheap lol. I added monthly free credits so people can test it, but once those run out you either wait for the reset or pay as you go. (LLM API costs still fall on you.)<p>Also: it’s very much a beta, so be careful with extremely complex tasks.<p>I’m considering hosting something like qwen3.5:122b so people can test without their own API keys. Credits would run out faster but it might be cheaper overall. Curious if people would want that.<p>Site:
<a href="https://usegradient.dev" rel="nofollow">https://usegradient.dev</a><p>Originally I planned to limit testing to ~20 people, but honestly I'm curious how it behaves with more users.<p>Pretty please don't DDOS me.<p>Any feedback is appreciated.