DoorDash's AI Assistant Combines LLMs with Specialized Agents and Memory Systems

DoorDash has released technical details about Ask DoorDash, a conversational AI system that helps users discover restaurants, organize meals, and shop for groceries through natural language. Rather than depending on a large language model alone, the platform combines specialized AI agents, Model Context Protocol tooling, consumer memory systems, and automated quality evaluation to deliver reliable experiences at scale.

Production Results Highlight Real-World Impact#

Early tests demonstrated tangible improvements across core business metrics. When handling grocery orders, the system's memory capabilities boosted checkout conversion by roughly 24%, pushed average basket sizes up 17%, and cut the number of back-and-forth exchanges by 7% during a week-long trial. For restaurant recommendations, the assistant drove a 15% jump in conversion rates when handling open-ended user queries. The company also accelerated its ability to measure and validate AI agent quality through automation, reducing the time needed to test model updates and system changes.

Architecture Separates Concerns Across Layers#

The design keeps orchestration logic separate from business capabilities. An Assistant Runtime manages interactions between different specialized agents, while a shared Model Context Protocol layer exposes DoorDash's existing services—catalog search, suggestions, shopping carts, checkout flows, order history, and user memory. This modularity allows the company to evolve backend systems independently while letting multiple AI experiences tap the same integrations. Rather than hard-coding business logic into prompts, the assistant calls reusable tools that connect to established DoorDash infrastructure.

Three-Layer Memory System Provides Context#

Personalization runs through a three-tier memory architecture. Long-term memory, built offline from past shopping and dining history, captures stable preferences like favorite cuisines and dietary needs. Session memory preserves the flow of a current conversation, while agentic memory records facts users explicitly share. The system retrieves relevant memories via semantic search, ranks them, and weaves them into the model's input without embedding memory logic in model inference itself. As DoorDash noted, "Agents do not just need access to user data. They need the right context, for the right task, at the right moment."