DoorDash's engineering team details the agent memory system powering Ask DoorDash, their agentic ordering assistant. The system bridges offline long-term memory (behavioral preferences, brand affinities, dietary patterns) with live agents through a three-layer architecture: offline memory generation, distributed SQL+vector storage, and an orchestration tooling layer. Key mechanisms include task-aware query planning across semantic and keyword search, multi-tenant namespace partitioning, and a conversational memory flywheel that asynchronously extracts durable facts from chat sessions and writes them back to long-term storage. The system applies domain-aware durability rules (e.g., 'I don't want ramen' is transient in restaurant context but brand preferences in grocery are captured immediately), two-stage semantic deletion to avoid over-deleting adjacent memories, and fact-lifecycle management based on what a fact is about rather than uniform TTL decay. Early production results show grocery sessions with memory converted to checkout at ~24% higher rates, basket size increased ~17%, and intent misunderstandings dropped ~33% compared to sessions without memory.
