Why Stateless LLMs Fail at Personalization

▶ Watch (00:08)

Without memory, a capable LLM still returns useless answers. The speaker asked a plain question: recommend places serving beverages near the ski lift. The response named Colorado and Whistler. Generic, because the model had no idea the user lives in Seattle, skis Whistler annually, and means coffee by “beverages.”

“So no memory, no rag anything else.”

That 30-second baseline run makes the gap concrete. A stateless LLM cannot personalize, no matter how capable the underlying model is.

▶ Watch (05:29)

Mem0 stores past conversations as structured memories in an Azure AI Search index. The speaker loaded three prior interactions: a preference note

“I enjoy skiing and outdoor activities.”

a chat history about annual Whistler trips, and an injury note flagging advanced and black slopes as off-limits. When the same beverage question ran again, the agent retrieved those memories first and passed them alongside the query to the LLM. The response shifted to Whistler gondola coffee spots. The model inferred that “beverages” meant coffee because the memory said so. See the memory-enabled travel assistant demo (08:28) for the before-and-after output side by side.

Multi-Agent Orchestration with AutoGen and Private Catalog RAG

▶ Watch (10:11)

Adding a product advisor as a second agent changes the shape of the system. AutoGen manages both: the travel assistant handles itinerary and slope questions, the product advisor handles gear and clothing. Each runs as a function tool the AutoGen orchestrator calls as needed. The product advisor combines memory retrieval with a live search over a private product catalog indexed in Azure AI Search. In the AutoGen dual-agent live run (13:51), the first query returned priced ski gear from the catalog. The second, asking about slopes without mentioning skill level, returned beginner recommendations because the injury memory was still active.

Notable Quotes

So no memory, no rag anything else. ▶ 4:31

I enjoy skiing and outdoor activities. ▶ 6:26

I also like drinking coffee with a Mountain View. ▶ 6:28

Key Takeaways

  • Mem0 stores past conversations as vector memories in Azure AI Search, enabling per-user personalization at scale.
  • The same query returns different results once memory is retrieved and passed alongside the prompt to the LLM.
  • AutoGen routes queries across specialized agents, letting memory and private catalog RAG work together without overlap.