What Building SupportMind Taught Us About AI Agents

SupportMind, an AI customer-support agent built during a hackathon, taught a team of engineers several valuable lessons about the capabilities and challenges of AI agents. The project began with a straightforward idea: create an AI agent that remembers customers. However, as the team delved deeper, questions arose about what the agent should remember, how to retrieve the right memory, and how to prevent customer histories from mixing.

Lesson 1 emphasized that an LLM (Language Model) and an agent are not the same thing. While an LLM can generate responses, the application around the LLM can decide what information it sees, which tools it can use, what it stores, and what happens after it responds. In SupportMind, the language model handled the conversation while the application managed customer-specific memory. This separation allowed the team to think about the system more clearly.

Lesson 2 highlighted that more context isn't always the goal. Initially, the team considered sending the entire customer conversation history to the model. However, for long-term support, it became clear that the goal was to provide the model with useful context, not simply more context. SupportMind only recalls relevant memories based on the current message.

Lesson 3 pointed out that identity matters. Long-term memory could become problematic if memories aren't separated correctly. If Priya's WiFi history appeared in Ramesh's support conversation, the feature would become a problem rather than a solution. To address this, the team used a customer identifier as the memory-bank identifier, implementing the function hindsight.recall(bank_id=customer_id, query=message).

Lesson 4 stressed the importance of testing the zero memory case first. The team naturally wanted to demonstrate returning customers, as this is where the project looked impressive. However, every returning customer was once a new customer. Testing with customers lacking relevant memory was crucial, as if nothing was recalled, SupportMind needed to tell the model that this was a new customer with no previous history.

Lesson 5 focused on making AI behavior observable. A particularly useful feature was displaying recalled memories beside the conversation. For example, if the assistant said, "The firmware update that solved your previous problem may be relevant again," the interface could show the previous memory responsible for that context. This made development and testing easier, as the team could inspect what information was being passed to the model.

Lesson 6 noted that summaries can be more useful than raw history. While long-term memory is valuable, support representatives may not want to inspect every individual memory. The customer briefing feature transformed previous interactions into a short summary containing important issues and successful fixes. This provided two ways of using memory: recalling information to answer the current question and reflecting to understand the customer more broadly.

The prototype stack for SupportMind was relatively simple, using Flask for the web application and API routes, Hindsight for customer memory, Groq with gpt-oss-120b for generating support responses, and a frontend to demonstrate various features such as customer selection, chat, recalled memories, comparison, and customer briefings.

Despite being a focused prototype, SupportMind answered a crucial question: what changes when an AI support agent can remember? The current limitations included the use of sample data and the inability of the assistant to access real customer accounts, issue refunds, modify subscriptions, or perform account actions. These limitations point toward the next stage of the project, where memory could be combined with authenticated customer accounts, ticket-management systems, CRM data, and carefully permissioned actions.

In conclusion, SupportMind demonstrated that giving past information to an AI support agent can significantly improve its responses. The project showed that the customer shouldn't start over; instead, their memory should be recalled and used to continue the conversation.