Original video in Chinese.
Key Takeaway
- “Global memory” is crucial for improving the coherence of AI conversations, enabling personal assistant functionality, and building global knowledge-sharing tools.
- The OpenMemory project achieves AI memory sharing across clients and across conversations by independently storing chat records and using the MCP protocol.
- OpenMemory’s functionality relies on large language models (for semantic understanding and retrieval), localized storage (to ensure privacy, data portability, and scalability), and the MCP protocol (to share memory across different clients).
- The author emphasizes the importance of context for AI to create value and points out that the “global memory” space offers opportunities for entrepreneurs to explore and grow.
- The suggestion is to separate knowledge base storage from invocation, store data locally, and call it through MCP, thereby gaining greater freedom in model choice.
- Intelligence is the means to attract users, while data is the key to retaining them over the long term.
Full Content
Although I often trash OpenAI, I very much agree with their judgment and views on “global memory.”
At a small scale, “global memory” is related to conversational coherence. This is one of the biggest pain points in AI experiences today.
Think about it: in most AI clients today, every time you start a new conversation, doesn’t it feel like you’re talking to a stranger? Because it doesn’t know what you talked about before, you have to explain everything again.
Doing it this way is inefficient, and it makes the tool feel very strong; you don’t feel like you’re talking to an intelligent being.
At a larger scale, “global memory” doesn’t just cover all historical conversations; it can and should also include all kinds of user data, and even a broader knowledge base.
Once the technology matures, it’s very possible to turn AI into a comprehensive personal assistant and a global knowledge-sharing tool. That is OpenAI’s ambition.
Of course, this vision is a bit far off. But standing here today, we can still experience it first—through the OpenMemory project.
Simply put, OpenMemory’s function is to store chat records separately. Anything discussed in any client can be saved. Then it can be invoked through MCP. Any client can call up any record. Let me show you, and it’ll make sense.
I’ve already installed OpenMemory and connected it to the ChatWise client via SSE. You can see that this MCP has four tools in total: add, query, retrieve, and delete.
I prepared a piece of text and asked the AI to save this content about Prompt House. After the LLM received the request, it called the add_memories tool through MCP and successfully saved that paragraph.
Then, in the same chat window, I asked: what content is in memory? You can see that the LLM smoothly pulled up the record I just saved.
Next, we open a new chat window and see whether the AI still “remembers” the previous content. The question is simple: what two problems does Prompt House mainly solve? Through the list_memories and search_memories tools, the LLM pulled up the content it had just stored and completed the answer based on it.
Within different chat windows of the same client, the LLM can freely store and invoke memory. So what about across apps? I opened Cursor and asked the same question: what two problems does Prompt House mainly solve? Claude Sonnet 4 also used the search_memories tool to retrieve the memory.
You see, whether across chat windows or across clients, OpenMemory MCP can store and retrieve memory.
To achieve these functions, OpenMemory does three things:
First, it calls a large language model.
To store a user’s conversations, the first thing you need to do is “understand” them—only by understanding the semantics can you distinguish noise from important information.
Not only that, when storing memory, you also need the large language model to identify conflicts within memory and perform updates. When retrieving, you also need the large language model to perform retrieval at the semantic level, not just keyword retrieval.
By default, OpenMemory uses OpenAI’s models, and we need to enter an OpenAI API Key. But it also supports models from other providers, including open-source large language models. If you care very, very much about privacy and security, you can connect Ollama.
Second, localized storage.
To put users at ease, OpenMemory stores all information locally. It uses a dual-database architecture, one for storing memory vector embeddings and one for storing structured metadata.
The advantages of this design are: first, it is fully local, so sensitive information is not uploaded to the cloud; second, the data is portable, and the entire data directory can be easily backed up and migrated; third, it is highly extensible, allowing you to switch between different large model providers or use local models.
Third, MCP.
Any client that supports the MCP protocol can connect to OpenMemory via SSE. Its MCP provides four tools: add, retrieve, view, and delete.
With these four tools, different clients can share the same memory. Context can be passed seamlessly between different clients.
I’m very optimistic about the direction of “global memory.” Although the giants are building it, entrepreneurs still have a lot of opportunities.
Because if it’s “global,” then it can’t be controlled by one company alone; it has to be open. In the current stage, which company can do that? Doesn’t that create a window for entrepreneurs to survive and grow?
Moreover, the giants are still mainly focusing their energy on improving model capabilities. So a few months ago, I shared an idea in the community:
Separate knowledge base storage from invocation. For storage, put it locally and use Cursor to create a Milvus vector database. For invocation, use Milvus MCP. At the time, everyone seemed quite interested, so I even made a community-exclusive video to introduce how to set it up.
From the user’s perspective, the advantage of this approach is that model choice becomes much more flexible. Whoever has the strongest model, I’ll connect to them.
From a product perspective, what ultimately retains users is not intelligence, but data. Intelligence is just a hook to attract users and complete the migration of data.
And intelligence, in order to work and create value, must rely on context. This is what I said before in the community: Context Matters.
Looking back, all those scattered ideas I shared over the past few months, along with my recent attempts, now all connect together.
OK, enough said. I’m going back to building products. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. All my insights and understanding will be shared in the community. See you next time!