请问HN:在物理代理之间进行任务规划时,选择RAG还是共享内存?
基于大型语言模型(LLM)的软件代理在工具使用、记忆和多步骤任务规划方面已经变得相当出色。但我很好奇是否有人在将这一技术进一步应用于物理世界,特别是与机器人或配备传感器的代理相关的领域。
例如:
想象一下,机器人A观察到某个物品位于区域Z,而机器人B随后需要去取回它。它们是如何共享这一上下文的?是通过:
- 一个结构化的记忆层(比如知识图谱)?
- 在一个基于RAG的集中状态存储中?
- 还是某种更简单(或更混乱)的方式?
我正在尝试使用共享知识图谱作为代理之间的记忆——通过RAG支持非结构化输入,并且可以查询以进行规划、依赖关系和任务调度。
我很想知道:
- 是否还有其他人考虑在物理代理之间实现共享记忆?
- 你们是如何处理世界状态、任务上下文或协调的?
- 有没有你们觉得有帮助的框架或经验教训?
我正在探索这个领域,非常希望听到其他在这一领域或周边进行建设的人的意见。
谢谢!
查看原文
LLM-based software agents are getting pretty good at tool use, memory, and multi-step task planning. But I’m curious if anyone is pushing this further into the physical world; specifically with robots or sensor-equipped agents.<p>For example:<p>Imagine Robot A observes that an item is in Zone Z, and Robot B later needs to retrieve it. How do they share that context? Is it via:<p><pre><code> - A structured memory layer (like a knowledge graph)?
- Centralized state in a RAG-backed store?
- Something simpler (or messier)?
</code></pre>
I’m experimenting with using a shared knowledge graph as memory across agents—backed by RAG for unstructured input, and queryable for planning, dependencies, and task dispatch.<p>Would love to know:<p><pre><code> - Is anyone else thinking about shared memory across physical agents?
- How are you handling world state, task context, or coordination?
- Any frameworks or lessons you’ve found helpful?
</code></pre>
Exploring this space and would really appreciate hearing from others who are building in or around it.<p>Thanks!