How to Properly Implement Memory for an AI Agent in Make?

How To
1

Hi everyone!
I’m working on creating a customer support agent for a WordPress site. I’m currently facing a challenge with memory handling.

The agent sometimes fails to respond properly when a customer’s message is ambiguous, because it lacks real memory of previous interactions.
To solve this, I saved the conversation history into a database and created a Make scenario that queries the conversation logs and provides the past context as a tool to the agent.

Problem:
Even though I query the conversation history and attach it to the request (as a tool or additional context), the agent still tends to skip or ignore this information, acting like it has no memory.

My questions are:

  • What is the best practice to inject past conversation history so the agent can truly use it?
  • Is there a way to make the agent read and incorporate that memory properly into its thinking, especially when using Make with OpenAI or similar tools?
  • Should I prepend the history into the system prompt, the user message, or manage it another way?
  • Is there a better way to create true memory for agents inside Make?

Thanks in advance for any tips or experience you can share!

2

Welcome to the Make community!

For more information, do view these similar questions previously asked by other community members, which may answer your question:

Next time, please try to search this forum first for existing answers. This may help you get to the solution faster, and also help keep the forum organised. Thanks for your cooperation!

Hope this helps! Let me know if there are any further questions or issues.

— @samliew

3

Hi Brian! Welcome to the Community!

Make AI Agents already includes a short-term memory capability, especially designed for conversation history.

It’s implemented through the thread parameter. That can be any value, but you usually relate it to some concept in the messaging platform you’re using - for example, the Slack timestamp parameter.

There’s a reference to that in this use-case from the Make website.

1 Like