展示HN：Neurotrace – 我开发但从未使用过的扩展程序
我在机器学习软件领域工作了四年，期间我遇到了一个反复出现的问题：我总是问自己，“我为什么要这样写这个函数？”或者“这个代码块为什么在这里？”
我尝试使用Obsidian和其他笔记应用来整理我的思路，但说实话，为自己写文档总是感觉像是一项繁琐的工作。文档总是让人觉得是为“别人”准备的。
于是，我决定构建一个VS Code扩展，直接将我的推理和上下文记忆与代码片段、标签等链接在一起。我甚至添加了一个优先级任务列表，这样我就能准确知道第二天待办的事情。
结果是什么？我从来没有使用它。
几个月后，我感到失望。我觉得自己在一件连我自己都觉得没用的事情上浪费了时间。
然后，智能代理出现了。
与AI代理的合作让我开阔了视野，但我遇到了一个瓶颈：“冷启动”问题。每次新会话都需要我从头开始解释一切。我尝试了MEMORIES.md、AGENTS.md和Claude的项目规则。讽刺的是，冷启动的改善并没有如预期那样明显。一些基准测试甚至显示，当被迫解析过多静态技能文件时，代理的表现反而更糟，而其他测试仅显示出边际10%的改善。
出于好奇，我决定实现一个本地的MCP，以便我的代理能够自主使用Neurotrace。
结果令人震惊。我没想到代理们真的会使用这个工具，但他们确实在使用。我还没有正式的基准测试，但我可以自信地说，冷启动的问题大大减少了。由于我使用来自不同提供商的不同代理，现在“下一个”代理确切知道我们昨天停在哪里了。他们决定保存哪些上下文记忆，而且做得相当不错。我的工作流程显著改善。
我很想听听你对代理记忆的看法，或者你是否找到更好的方法来处理上下文交接。
祝好，
Irwing Castro (CastleOneX)
你可以在以下市场找到它：
VS Code市场: [https://marketplace.visualstudio.com/items?itemName=BlackIronTechnologies.neurotrace](https://marketplace.visualstudio.com/items?itemName=BlackIronTechnologies.neurotrace)
Open VSX: [https://open-vsx.org/extension/BlackIronTechnologies/neurotrace](https://open-vsx.org/extension/BlackIronTechnologies/neurotrace)
查看原文
I’ve been working in ML software for 4 years, and I quickly ran into a recurring problem: I kept asking myself, "Why did I write this function this way?" or "Why is this block here?"<p>I tried to organize my thoughts with Obsidian and other note-taking apps, but let’s be honest, documenting for yourself feels like a chore. Documentation always feels like it's meant for "someone else."<p>So, I decided to build a VS Code extension to save my reasoning and contextual memory directly linked to snippets, tags, and more. I even added a prioritized task list so I’d know exactly what was pending the next day.<p>And what happened? I never used it.<p>Months later, I felt disappointed. I felt like I had wasted all that time on something even I didn't find useful.<p>Then, agents arrived.<p>Working with AI agents has been a mind-opening experience, but I hit a wall: the "Cold Start" problem. Every new session required me to explain everything from scratch. I tried MEMORIES.md, AGENTS.md, and Claude’s project rules. Ironically, the cold start didn't improve as much as promised. Some benchmarks even show that agents perform worse when forced to parse too many static skill files, while others only show a marginal 10% improvement.<p>Out of curiosity, I decided to implement a local MCP so my agents could use Neurotrace autonomously.<p>The result was startling. I didn't think the agents would actually use the tool, but they are. I don't have formal benchmarks yet, but I can say with confidence that the cold start has drastically decreased. Since I use different agents from different providers, the "next" agent now knows exactly where we left off yesterday. They decide what contextual memories to save, and they do it surprisingly well. My workflow has improved significantly.<p>I'd love to hear your thoughts on agentic memory or if you've found better ways to handle context hand-off.<p>Cheers,<p>Irwing Castro (CastleOneX)<p>You can find it on the marketplaces here:<p><pre><code> VS Code Marketplace: https://marketplace.visualstudio.com/items?itemName=BlackIronTechnologies.neurotrace
Open VSX: https://open-vsx.org/extension/BlackIronTechnologies/neurotrace</code></pre>