Hello guys, I’m building an AI agent for real estate companies. I’m almost done, but I saw someone on YouTube saying that simple memory is just for experiments, not for real-world use.I don’t have a problem with simple memory, but I don’t know how it will be if 100-200 people talk to the agent. My goal with simple memory is just conversation, because I’m storing the data in Airtable like customer information and interests.
More details:
This is my AI agent that I am building. I use WhatsApp.Now I’m using simple memory in all agents with the same ID. So they all share the same memory. One of the tasks of the Router agent is to save the name, phone number, and interests of the customers in Airtable. Now what am I supposed to change so that I can sell with confidence?.
Hi @Ali4 ,
You can try this to make it better: Instead of using simple memory, switch to a session-based memory system like MongoDB, Redis, or Postgres. This will keep each person’s chat separate by using their WhatsApp number or user ID as a unique session. You can still save customer details in Airtable just like you’re doing now. This way, your agent will handle more users smoothly
Thanks for the reply. I’ve modified the system as per your advice and I have two additional questions that I hope you can help me with. I mentioned that I use the Router Agent to save customer data, but I face a problem with duplicating data, for example, two Rows containing the name, number, and customer interest, but the first Row is missing the customer interest. This means that the AI agent saves the data even if not all the data is there, and when it collects all the data, it saves it again. Does the prompt solve this problem, or is there a better way to save customer data?As for my second question, what is the best LLM for AI clients? I use Gemini 2.5 and I found it better than GPT-4.1 mini in terms of using tools and responses, but at the same time, Gemini is not the best either. Especially in the use of tools. Does the prompt solve this problem also?
By the way, I visited your website a while ago and I was impressed with your work.
Regarding the first question, I think the database processing logic should be optimized so that database write operations are only executed by a single node. Regarding the second question, my experience is that there is no single model currently that is particularly suitable for the AI client. After testing the tool call boundaries of the model, it is recommended to split complex calls and handle them in the form of sub-nodes or sub-models.
Thanks for the feedback — and we’re glad you liked our work!
On the duplicate data issue, prompting is a big part of making the AI agent behave correctly, so you’ll want a good prompt with clear conditions. For example: “Only save customer data if all required fields (name, number, interest) are present. If the record exists, update it instead of creating a new one.” We also recommend adding a check in your flow to detect duplicates before saving — this way you have both the AI logic and the workflow safeguard working together, so partial info won’t create extra rows.
For models, here’s a quick comparison from our experience: GPT-4.1 – strong reasoning & tool use, higher cost; GPT-4.1 mini – cheaper, faster, slightly less accurate; Gemini 2.5 – great tool use, often cheaper than GPT-4.1, but less consistent in complex logic; Claude 3.5 Sonnet – excellent for structured outputs & long context, mid-range cost. For the best balance of cost vs. results, GPT-4.1 mini or Claude 3.5 are solid. For heavy reasoning, GPT-4.1 is still the leader.
If you have any other problem, you’re free to ask.
To store your data in Airtable without duplicates, you should first collect all the required fields like name, phone number, and interest before saving anything. Use temporary storage in your workflow to hold the data until it’s complete. Then, before saving, check Airtable using the phone number to see if the customer already exists. If they do, update the existing record; if not, create a new one. This way, you avoid saving incomplete or repeated data. You can do this using Airtable’s API
I was late in responding because I was adjusting the system. Now everything is fine. I took my time testing the models(GPT-4mini, GPT-5mini, and Gemini 2.5 flash). Gemini is much better at formulating responses and arranging words. GPT-5 Mini is slow and philosophizes a lot in responses, but it may be better at using tools. However, Gemini isn’t bad at this either. I had GPT-5 help me shorten and enhance the prompt, and it’s good at that. Now everything is fine, thank you, and sorry for the inconvenience.
