AI-Driven User Memory System for Dynamic Real Estate Personalization

zillow · Tech · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Personalized real-estate discovery

Zillow needed personalization that could remember long-running home-shopping intent while adapting when buyers changed cities, property types, budgets, or soft preferences.

Constraints:

months-long decision journeys
fast-changing user intent
privacy and control requirements
multi-surface search, recommendation, and notification experiences

Architecture

Zillow models personalization as a layered user-memory system combining raw interaction replay, batch and streaming preference pipelines, affordability models, embeddings, and privacy controls that feed search, recommendations, and notifications.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  raw-interaction-history["Raw interaction history"]
  user-memory-layer["User memory layer"]
  batch-profile-pipelines["Batch profile pipelines"]
  streaming-behavior-pipelines["Streaming behavior pipelines"]
  preference-profiles["Preference profiles"]
  affordability-quantile-models["Affordability quantile models"]
  deep-user-embeddings["Deep user embeddings"]
  privacy-controls["Privacy controls"]
  search-and-recommendation-surfaces["Search and recommendation surfaces"]
  raw-interaction-history --> user-memory-layer
  batch-profile-pipelines --> preference-profiles
  streaming-behavior-pipelines --> preference-profiles
  preference-profiles --> search-and-recommendation-surfaces
  affordability-quantile-models --> search-and-recommendation-surfaces
  deep-user-embeddings --> search-and-recommendation-surfaces
  privacy-controls --> user-memory-layer
  user-memory-layer --> search-and-recommendation-surfaces

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Raw interaction history`	Preserves exact user actions when replayable continuity matters.	`User memory layer`
`User memory layer`	Combines raw history, batch profiles, streaming signals, affordability, embeddings, and privacy constraints into reusable personalization memory.	`Raw interaction history`, `Batch profile pipelines`, `Streaming behavior pipelines`, `Preference profiles`, `Privacy controls`, `Search and recommendation surfaces`
`Batch profile pipelines`	Model long-term preferences over extended observation windows.	`Preference profiles`, `User memory layer`
`Streaming behavior pipelines`	Capture recent behavior and intent changes for responsive personalization.	`Preference profiles`, `User memory layer`
`Preference profiles`	Represent recency-aware preferences, priorities, and intent.	`Search and recommendation surfaces`
`Affordability quantile models`	Estimate personalized affordability constraints and price sensitivity.	`Search and recommendation surfaces`
`Deep user embeddings`	Generalize user interests across listings and surfaces.	`Search and recommendation surfaces`
`Privacy controls`	Constrain memory use around user rights, control, and data handling.	`User memory layer`
`Search and recommendation surfaces`	Apply memory to search, recommendations, notifications, and personalization.	`User memory layer`

Data flow

Behavioral signals flow into batch and streaming profile pipelines, become recency-aware preferences, affordability estimates, embeddings, and raw-history memory, then power personalized Zillow experiences.

“it’s a layered, AI-driven profile of preferences, priorities and intent — all evolving over time. At Zillow, we’re building this memory across three dimensions”article, stated

“From recency-aware preference profiles to affordability-aware quantile models and deep user embeddings, these components form the foundation of a memory layer that powers personalized, responsive and consistent experiences.”article, stated

Why this architecture

The memory architecture exists because long home-shopping journeys include both stable preferences and sudden intent shifts, and inferred preferences alone cannot cover every continuity need.

“Capturing the right level of nuance (what’s a strong signal vs. a soft preference) * Adapting gracefully to changes in behavior and intent * Preserving helpful context”article, stated

“While models help us infer and generalize user preferences, there are moments where the most useful memory is simply an accurate replay of what happened. That’s where raw interaction history comes in.”article, stated

Models used: embedding models · quantile models · preference modeling systems

Scale

The personalization system includes real-time preference modeling, affordability estimation, and user embeddings.
“This memory is shaping search, recommendations and other personalized experiences on Zillow through real-time preference modeling, affordability estimations and user embeddings.”article, stated

Alternatives considered (roads not taken)

Only short-term real-time signals or only long-term batch profiles

Rejected because: Zillow uses both because batch pipelines capture extended preference windows while real-time behavior catches fast intent changes.

“Batch user-profile pipelines are well suited for modeling long-term preferences because they operate over extended observation windows and can employ higher-latency, more complex algorithms.”article, stated

Only model-inferred preferences

Rejected because: Zillow keeps raw interaction history because some experiences need exact replay of what happened.

Personalization without explicit privacy controls

Rejected because: The source says tailored experiences should not come at the expense of privacy, control, or data rights.

“giving consumers a more tailored experience should never come at the expense of their privacy, control or data rights. That’s why our personalization infrastructure is designed to”article, stated

Patterns

`agent-memory` (context)

The system treats personalization as a living user memory built from explicit preferences, inferred intent, embeddings, affordability models, and raw behavioral history.

“it’s a layered, AI-driven profile of preferences, priorities and intent — all evolving over time. At Zillow, we’re building this memory across three dimensions”article, stated

`privacy-preserving-processing` (safety)

Personalization infrastructure is constrained by privacy, data-rights, identity protection, and user-control requirements rather than only model quality.

Tradeoffs

How to represent long-term and short-term intent in the same memory system.

Chose batch jobs for slow-moving profiles plus streaming pipelines for transient behavior over one latency profile for all personalization signals.

Rationale: Daily batch processing can handle complex long-term modeling efficiently, while streaming is needed for rapid shifts in short-term behavior.

Gotchas (what bit them)

Stale user assumptions can make personalization feel wrong during a long home-search journey.

Root cause: Home shoppers change priorities over weeks or months, so a static profile can overfit past behavior.

Resolution: Model memory across recency/frequency, flexibility/rigidity, and prediction/planning so the system can distinguish durable preferences from temporary signals.

“Capturing the right level of nuance (what’s a strong signal vs. a soft preference) * Adapting gracefully to changes in behavior and intent * Preserving helpful context”article, implied

Inferred preferences are not enough for experiences that require user-visible continuity.

Root cause: Models can generalize from behavior, but some product surfaces need an accurate replay of what the user actually did.

Resolution: Retain raw timestamped interaction sequences alongside model-derived profiles.

Evaluation

The article describes product surfaces powered by the memory layer but does not provide offline evaluation or production metric methodology.

“This memory is shaping search, recommendations and other personalized experiences on Zillow through real-time preference modeling, affordability estimations and user embeddings.”article, stated

Outcomes

Zillow reports a memory layer that supports personalized search, recommendations, notifications, and continuity across sessions.
The architecture is more ML personalization than explicit LLM agent orchestration.

Zillow engineering article; no quantitative lift or reliability metrics are provided.