DoorDash built a unified consumer memory platform that extracts semantic understanding from behavioral signals (orders, browsing, searches) and makes it available across all personalization surfaces. The system uses LLMs to generate structured 'memory blocks' — versioned, domain-specific summaries like dietary preferences, brand affinities, and dining patterns — organized into three layers: long-term, in-session, and explicit context. These memory blocks are encoded two ways for ML consumption: dense asymmetric embeddings for semantic similarity matching, and a heterogeneous context graph for relational reasoning between consumers, brands, taxonomies, and concepts. The platform powers personalized carousels and ranking models, with selective recomputation to manage LLM compute costs. Key lessons include treating memory blocks as semantic matching primitives for both ML and LLM systems, decoupling extraction from encoding, and maintaining full versioning/lineage for every component.
