Building a meeting-prep agent with real memory in ~60 lines: Hindsight + Groq

UmeshGupta2008 · reddit · 2026-11-07

A first-year AI/ML student shares how they built a meeting prep & action-item agent in 60 lines of code, addressing LLMs' statelessness.

Core idea: stop stuffing context, start recalling it

Instead of dumping all past notes into the prompt (expensive, slow, noisy), split into two verbs:
Retain: store notes in Hindsight's long-term memory;
Recall: fetch only relevant memories when asked.
Architecture is deliberately boring: input → router → [retain] or [recall + Groq inference] → response, all inside a while True loop; question detection uses simple keyword triggers (prep/what/who/when/remind/?).
No vector DB config, embedding pipeline, or chunking strategy—memory is Hindsight's job; reasoning runs on llama-3.3-70b-versatile via Groq.

In practice: after meetings, type notes like "John promised slides by Thursday"; before the next call, type "Prep me for my meeting with John" and the agent returns open commitments, relevant decisions, and a suggested opener.

The author's takeaway: memory turns a stateless chatbot into a collaborator that accumulates a relationship with your work.

Original post →

Building a meeting-prep agent with real memory in ~60 lines: Hindsight + Groq

More from coding & agent