展示HN：我的OpenClaw知道它一周前做了什么。感谢“hmem”-MCP。
类人记忆的人工智能代理
这就是我在床上躺着、试图入睡时思考的事情。
我厌倦了我的代理因为压缩记忆而忘记几天前告诉它们的事情。使用Gemini时情况更糟，有时上下文会被重置，它甚至会忘记五分钟前的事情。
于是我开始思考我如何组织我的记忆。我们并不会同时记住所有经历过的事情，而是将它们分块存储。当你想到“假期”时，你会想起自己去过的所有假期。然后你可以深入回忆某个特定的记忆，比如那次去罗马的旅行。再深入一点，你会想起酒店、床是否舒适、食物的味道等等。同时你也会想起和你一起去的人，当你想到他们时，会有更多与你的记忆相连的事情。那些你三分钟前还没想到的事情，但它们就在你的记忆中，随时可以调取。
然后我意识到，这个概念应该可以转化为一个MCP。第二天早上，我把这个想法告诉了Claude，我们做了一个原型。结果，它几乎完美地实现了。我们创建了可以正确使用MCP的技能，我该怎么说呢……有时我看到它在没有我要求的情况下使用它的hmem-MCP。它可以自主地读取和写入记忆。
想象一下吧。再也不需要低效的.md记忆文件，这些文件会淹没上下文并浪费你的令牌。
如果你想了解它是如何工作的，可以查看我的Github。
它仍在开发中，但已经运行得相当不错！如果你把Github链接传给你的Openclaw，它就能安装并使用它。
[https://github.com/Bumblebiber/hmem](https://github.com/Bumblebiber/hmem)
查看原文
Humanlike memory for AI Agents.<p>That's what I was thinking about when I lay in bed one more time while trying to fall asleep.<p>I was sick of my agents forgetting stuff that I told them a few days ago due to compressing their memories. With Gemini it was even worse, sometimes the context just got reset and it wouldn't remember things from 5 minutes ago.<p>So I thought about how I organize my memories. We don't have all the things we've experienced in mind at the same time. We keep them in junks. When you think about "holidays" you will remember all the holidays you've been to. Then you can dig deeper into one memory and think about that trip to Rome. Go even deeper and you'll remember the hotel, if the bed was comfortable, how the food was, etc. Also you'll remember the people who were there with you and when you think about those, there will be many more memories that connect with them. Things you hadn't thought about 3 minutes ago, but they were there in your memory. Just on demand.<p>Then I realized, this concept should be transformable to an MCP. Next morning, I told Claude about this idea and we made a prototype. Well, and it pretty much nailed it. We created skills to properly use the MCP and what should I say... I sometimes see it using it's hmem-MCP without me asking to do it. It reads and writes memories on its own.<p>Imagine that. No more inefficient .md-memory-files that will flood the context and waste your tokens.<p>Check my Github if you want to know how it works.<p>It's still in development but it already works pretty well! If you pass the Github-link to your Openclaw, it will be able to install and use it.<p><a href="https://github.com/Bumblebiber/hmem" rel="nofollow">https://github.com/Bumblebiber/hmem</a>