Building a Customer Support Agent with Persistent Memory

A customer support agent was created with the ability to retain and recall previous conversations, enabling it to provide more effective assistance to users. The key to this system is persistent memory, which is achieved through the use of a memory layer called Hindsight.

Conventional customer support agents function by taking a customer's message, feeding it into an LLM, and generating a response. This process works well for individual conversations, but it becomes problematic when the same customer returns later, as the agent has no context from previous interactions.

This new customer support agent aims to solve this issue by retaining important customer interactions as persistent memories. These memories can then be retrieved when needed, providing the agent with relevant context for subsequent interactions.

Hindsight, the architecture behind this system, offers three core operations: retain, recall, and reflect. The retain operation processes information and stores it in structured memories. Recall searches these memories for relevant information, and reflect synthesizes responses from these memories.

Useful memories for customer support include previous conversations, reported problems, products involved, troubleshooting steps, resolutions, customer preferences, recurring issues, and commitments made by support. For example, if a customer reports that their laptop battery drains quickly, the agent can retain information about previous troubleshooting steps and outcomes.

When a customer contacts the support system again with a similar issue, the agent retrieves the relevant history and can respond with context-specific information. This memory layer preserves facts and relationships rather than treating the entire conversation as one large text blob.

When a customer sends a new message, the agent first searches memory for relevant information using a query. It then combines the retrieved memories with the current request, creating a prompt that includes the relevant customer history. This prompt is then fed into an LLM, which generates a context-aware response.

By implementing persistent memory through Hindsight, this customer support agent can provide more personalized and efficient assistance, avoiding repetitive questions and improving overall user experience.