正在开发一种可在各个AI代理之间使用的Plug and Play个人AI记忆系统。
我一直在进行一个项目（Make0 AI），目标是为Codex和Claude Code提供一个共同的上下文。问题很简单——在使用Claude时，由于每天/每周的使用限制，我不得不在Codex中重新开始。当然，我可以让Claude总结一个我可以交给Codex的交接内容，但这效率不高。
因此，我想到在本地为它们创建一个共同的上下文，使用一个简单的指令列表，比如一个共享的md文件。但问题是，这个文件不断增长，最终共同的上下文超过了2万个token，导致token预算不必要地膨胀。
于是，我在这个列表上实施了一种基于RAG的检索方法，只加载该提示所需的内容，经过几次迭代，它最终形成了一个图形记忆——大脑。
这个产品Make0 AI使你能够创建大脑，并通过MCP连接到任何AI代理，这样它们就能立即了解你的喜好、厌恶、偏好等。AI代理可以读取和写入这个记忆。
这可以成为你的长期记忆，而代理仅仅充当token提供者和任务执行者。你可以从一个地方完全控制你的数据/记忆。可以把它想象成一个“USB闪存驱动器大脑”，你可以将其连接到任何代理。
我知道我可能不是第一个实现这个的人，但我尽量使用户体验简单，并且完全不以开发者为中心。你只需将你的记忆/数据拖放到大脑中，甚至可以与可视化的大脑进行互动。
还有一件事，我对记忆进行了基准测试。LoCoMo的整体得分为83.3%，在标准模式下单跳得分为92.5%（现已可用），在另一种配置（实验最大模式）下，整体得分为92%，单跳得分为94.4%。基准测试页面上可以查看。
目前我将其保持为邀请制。如果你们想试试，请告诉我！
查看原文
So I've been working on a project (Make0 AI) for a long time where I wanted to give common context to both Codex and Claude Code. The problem is simple - while working on Claude and after exhausting the daily/weekly limits I've start again from scratch in Codex. Of course I can prompt Claude to summarise a handoff that I can give to Codex but that is inefficient.<p>So I thought of a creating a common context for both of them locally using a simple list of instructions like a shared md file. But the problem was the file kept on growing and after some point the common context went beyond 20k tokens and was just inflating the token budget unnecessarily.<p>Then I implemented a RAG based retrieval approach on the list to only load what is required for that prompt and voila, after a few more iterations it resulted into a graph memory - Brain.<p>So this product Make0 AI gives you the ability to create Brains that you can connect to any Ai agent via MCP and then they immediately know your likes, dislikes, preferences etc. AI agents can read, write to this memory.<p>This can be your long term memory and the agents just become the token providers and task executors. You have full control over your data/memories from a single place. Think of this as a "USB thumb drive brain" that you plug to any agent.<p>I know I may not be the first one to implement this but I tried to make the UX simple and is not at all developer focused. You can just drag and drop your memories/data to the brain and even play around with a visual brain.<p>One more thing, i've done benchmarks on the memory. LoCoMo scores 83.3% overall, with Single hop 92.5% in standard mode (available now) and in another configuration (experimental max mode) it scores 92% overall and 94.4% in single hop. Available in the benchmark page.<p>I'm currently keeping this invite only. Do let me know if you folks wanna try it!