DoorDash Lifts Grocery Conversion 24% With Task-Aware Agent Memory

DoorDash's Head of New Verticals ML/AI, Sudeep Das, presented the company's production agentic recommendation architecture at QCon AI Boston on June 1, 2026. Core insight: one-shot ranking models don't compose into agents. DoorDash rebuilt three layers—catalog representation, consumer memory, and query intent classification—before moving from prediction to planning.

The catalog layer uses RQ-VAE semantic IDs: a residual-quantized variational autoencoder mapping product embeddings into discrete tokens instead of opaque integer item IDs. With 100 million+ SKUs across food, grocery, retail, and alcohol, this enables two capabilities. First, hierarchical RAG becomes feasible—the system narrows retrieval using category trees before the LLM sees a prompt, keeping inference fast and context windows compact. Second, LLM outputs map directly to meaningful product identifiers, eliminating the catalog-to-prediction resolution step that burns latency and token budget in naive RAG.

Memory is the second layer, where DoorDash reported clearest production gains. The agent memory system maintains three tiers: long-term memory from offline behavioral history (organized into versioned blocks—Dietary Preference, Dining Patterns, Item Brand, Item Taxonomy, Store Preferences, Cross-Channel Patterns); in-session context with real-time signals and recency weighting; and conversational memory from facts stated during a session. Sessions backed by this stack converted to grocery checkouts at 24% higher relative rate and restaurant-discovery queries at 15% higher. Intent misunderstanding dropped 33% on LLM-judge evals. Basket sizes increased 17% and conversational turns dropped 7% in a seven-day production run.

The engineering challenge isn't storing memory—it's serving it relevantly. Naive top-K retrieval wastes context: "find a dinner spot" requires cuisine tendencies and price history; "I want to make chicken tikka masala" requires pantry inventory and portion preferences; "shop my usuals" requires replenishment cadence. DoorDash solved this with task-aware query planning: the memory system reads the current task before retrieval, runs hybrid search (semantic vectors plus structured filters), ranks results, then injects into prompt. Conversational signals extract into durable memory blocks asynchronously.

The third layer addresses query intent classification. Multi-category marketplaces like DoorDash's—Food, Grocery, Retail, Alcohol on a single surface—break traditional softmax classifiers. Raising confidence in "Restaurant" for "Wildflower" suppresses "Retail," even when both apply. DoorDash built an Agentic Multi-Source Grounded system: an LLM grounded in staged catalog entity retrieval, augmented with agentic web-search for cold-start and long-tail queries. It emits an ordered multi-intent set with a deterministic disambiguation layer applying business policies. In production, this runs +10.9pp above an ungrounded LLM and +4.6pp above the prior hybrid BERT+LLM system. On long-tail queries, catalog grounding contributes +8.3pp, agentic web search adds +3.2pp, and dual-intent disambiguation adds +1.5pp, reaching 90.7% accuracy. The system serves 95% of daily search impressions via offline batch-and-cache—the LLM is not on the hot path.

Platform separation prevents complexity collapse. Domain teams own specialized agents (grocery, restaurant, Dasher-assist); a platform team owns orchestration, MCP tooling, memory, evaluation infrastructure, and shared components. Agent interfaces are identical across verticals but receive vertical-aware extraction rules, retrieval sources, and ranking signals. Deterministic actions—those that update versioned artifacts—skip the LLM entirely.

Task-aware context retrieval, not model scale, moved conversion across DoorDash's agent stack.