𝗜 𝗧𝗮𝘂𝗴𝗵𝘁 𝗛𝗶𝗻𝗱𝘀𝗶𝗴𝗵𝘁 𝘁𝗼 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿 𝗦𝗮𝗹𝗲𝘀 𝗖𝗮𝗹𝗹𝘀 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿𝗶𝗻𝗴 𝗡𝗼𝗶𝘀𝗲
Transcripts are too noisy to be memory.
I built a Deal Intelligence Agent to solve a specific problem. I did not want a long-term junk drawer of meeting transcripts. I wanted an agent that remembers what matters and forgets the filler.
I used Next.js 16, FastAPI, Supabase, Hindsight, and Groq to build this.
The system works by splitting data into two paths:
• Supabase stores exact facts. This includes deals, meetings, transcripts, objections, and action items. • Hindsight stores semantic memory. This answers: "What has this deal taught the agent so far?"
To keep the agent smart, I implemented three types of memory:
- Episodic memory: What happened in a specific meeting.
- Semantic memory: Patterns found across multiple meetings.
- Procedural memory: Strategies that worked for this specific deal.
I also added a logic gate to prevent the agent from being too confident too early.
The agent stays humble. One meeting only creates episodic memory. After two meetings, it starts to find patterns. After three meetings, it can suggest a winning strategy. This prevents the system from inventing patterns based on a single conversation.
The result is a massive difference in the pre-meeting brief.
A generic agent says: "Prepare to address budget concerns."
My agent says: "Sarah Chen, the CFO, rejected the enterprise price in meeting 2. She softened after the phased pricing proposal in meeting 4. Lead with the phased structure today."
The goal is not to replace the salesperson. The goal is to make the agent think like an experienced rep.
Lessons learned:
- Transcripts are source material, not memory. Extract facts first.
- Memory needs timing. Do not suggest strategies after just one meeting.
- Use different tools for different jobs. Use Supabase for facts and Hindsight for context.
- Names matter. Remembering "Sarah Chen" is better than remembering "the customer."
Optional learning community: https://t.me/GyaanSetuAi