No comments yet.
I know everyone is quite tired of the term "Memory System" recently. I was among the earliest developers who disliked this concept (maybe starting in August 2024? When mem0 popularized the concept). Back then, I felt that chatbots/agents were originally clean state machines, and everything became chaotic after introducing a memory layer.
However, I later realized that although the term "Memory" seems to be a very general solution, it actually needs to be somewhat related to the business scenario. What first changed my mind was when ChatGPT officially emphasized its memory in 2025 [0]: I observed that so many users felt delighted because of the memory.
At that time, I thought that perhaps for consumer-facing chatbots aimed at consumption, entertainment, and life scenarios, memory might be an important feature after all. In this context, I believe the memory system should meet the following requirements:
- Low online latency: Consumers will rarely choose a product because of its "memory" capability, but they will immediately abandon it if the first-token response time takes too long. Memory should enhance user experience, not eat into the initial response time. It was not supposed to be the core of a product anyway.
- Modeling user instead of searching data: For consumer applications, 99% of queries are not asking AI "what did you say before?" People want proactive associations, not passive search results. Instead of spending efforts on conversation RAG, it might be more effective to refer to the bio tool[1] in ChatGPT and maintain a global user profile of 500-1000 tokens.
This is also why I appreciate the design of ChatGPT Memory: They know what they aim for. You don't notice much latency, yet it can answer questions like "What kind of person do you think I am?" -something search alone can't solve. That's when I realized that memory for consumer AI should be quite different from those open-source memory systems filled with "embedding," "graph," "weight decay," "temporal relationship," etc. A simple approach can be enough.
This is also why I built Memobase:
- Built with pure Python: easy to read and modify. Only depends on Postgres and Redis
- User Profile/Timeline: Each user has an automatically modeled and updated profile (kept under ~1000 tokens) plus a chronological event timeline.
- Highly customizable: You can define any profile dimensions and preferences of memories, rather than letting AI decide what should or shouldn't be remembered.
- Context API: Memobase will directly return the user's personal context as a prompt string, which you can directly insert into your system prompt as a supplement to user information. By default params, its latency is below 100ms.
We've also run some benchmarks. For example, on the LOCOMO dataset, Memobase achieved an overall accuracy of about 74-75%[2] (vs. mem0 68%[3]). However, I also want to point out that such benchmarks are not suitable for testing whether a memory system can enhance your product's user experience. In most public test results, ChatGPT Memory scores are frustratingly low, yet in reality, it is the only memory to have a widespread impact on users.
This is also why I am confident in Memobase: its underlying structure is so simple, customizable and fast. It is a solution I developed after observing what works in ChatGPT memory. If you also want to create a consumer-facing AI product similar to ChatGPT, I hope you can check out Memobase on GitHub.
Check it out on GitHub: https://github.com/memodb-io/memobase
[0]: https://x.com/sama/status/1910380643772665873
[1]: https://www.reddit.com/r/ChatGPT/comments/1fzq4uc/what_is_bi...
[2]: https://github.com/memodb-io/memobase/tree/main/docs/experim...
