Multi-Agent Personalization Engine with Proactive Memory System for Batch Processing

personize.ai · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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: Generative personalization

Personize.ai needed autonomous batch personalization over customer databases without inconsistent agents, context overload, shallow customer understanding, or ongoing human intervention.

Constraints:

The system operates over large CRM/customer databases in batch mode.
There is no human in the loop during execution, only high-level supervision.
Traditional RAG and function calling led to inconsistency and context overload.
Business accuracy must be defined per customer company and domain.

Architecture

Personize.ai built Cortex, a multi-agent personalization engine with a proactive memory system. Agents perform research, inference, planning, and content generation, while a centralized memory layer infers and synthesizes customer insights into standardized attributes shared across agents. This compresses context and enables consistent personalization for emails, websites, and product pages.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  customer-database-interface["Customer Database Interface"]
  research-agents["Research Agents"]
  proactive-memory-system["Proactive Memory System"]
  planning-agents["Planning Agents"]
  content-generation-agents["Content Generation Agents"]
  customer-database-interface --> research-agents
  research-agents --> proactive-memory-system
  proactive-memory-system --> planning-agents
  proactive-memory-system --> content-generation-agents
  planning-agents --> content-generation-agents

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

Components

Component	Role	Interacts with
`Customer Database Interface`	Connects Cortex to CRM and customer data for batch processing.	`Research Agents`
`Research Agents`	Perform deep customer and account research.	`Proactive Memory System`
`Proactive Memory System`	Synthesizes customer insights into standardized attributes shared across agents.	`Planning Agents`, `Content Generation Agents`
`Planning Agents`	Plan personalized content strategy from memory attributes.	`Content Generation Agents`
`Content Generation Agents`	Generate personalized emails, webpages, and product-page content.	—

Data flow

Cortex reads customer database records, performs deep research and inference, writes standardized customer attributes into proactive memory, then downstream agents recall that shared memory during planning and content generation for personalized emails, websites, or product pages.

“Their pipeline includes deep research capabilities, inference and reasoning operations, planning phases, and finally content generation for individual customers.”summary, stated

“Their solution implements a proactive memory system that infers and synthesizes customer insights into standardized attributes shared across all agents”summary, stated

Why this architecture

Personize.ai needed a memory-centric architecture because autonomous personalization over databases requires more than attractive generated prose; agents need a shared, compressed understanding of each customer that remains consistent across independent builders and content surfaces.

“effective personalization requires more than simply generating beautiful text—it demands a systematic, agentic approach with multiple sequential steps.”summary, stated

“different agents may be built by different teams or builders, yet they must generate consistent output across the entire system when operating at scale.”summary, stated

Models used: unspecified multi-agent LLM stack

Scale

The system had been tested with more than 20 B2B companies.
“Early testing with 20+ B2B companies showed the system can perform deep research in 5-10 minutes”summary, stated
Deep research reportedly takes 5-10 minutes.
“the system can perform deep research in 5-10 minutes and generate highly personalized, domain-specific content”summary, stated
The work had been tested over a three-month period.
“tested with over 20 B2B companies over a three-month period”summary, stated

Alternatives considered (roads not taken)

Use traditional RAG and function calling directly

Rejected because: Those approaches caused inconsistency, context overload, and weak customer understanding.

“traditional RAG and function calling approaches when processing customer databases autonomously, including inconsistency across agents, context overload”summary, stated

Use static templates and token substitution

Rejected because: Traditional personalization lacked nuance and senior-level quality.

“traditional personalization approaches, which rely on static templates and token-based substitution.”summary, stated

Patterns

`agent-memory` (state)

A proactive memory system centralizes customer insights as standardized attributes for all agents.

“standardized attributes shared across all agents, enabling centralized recall and compressed context.”summary, stated

`multi-agent-orchestration` (architecture)

Research, inference, planning, and generation agents run as a multi-step personalization pipeline.

“Their pipeline includes deep research capabilities, inference and reasoning operations, planning phases, and finally content generation”summary, stated

`context-engineering` (context)

The system compresses customer context into reusable memory attributes to prevent context overload.

“enabling centralized recall and compressed context.”summary, stated

Tradeoffs

Use proactive shared memory

Chose Standardized customer attributes shared across agents over Each agent independently retrieving raw context.

Rationale: Shared memory reduces inconsistency and context overload.

“including inconsistency across agents, context overload, and lack of deep customer understanding.”summary, stated

Run without per-item human review

Chose Autonomous batch execution with high-level supervision over Human-in-the-loop review for every generated item.

Rationale: Batch personalization at scale would not work if every item needed human intervention.

“There is no human in the loop during execution, only high-level human supervision overseeing the system.”summary, stated

Gotchas (what bit them)

Good-looking personalized text can still be inaccurate.

Root cause: Early client feedback showed generated copy did not always convey meaningful or accurate customer information.

Resolution: Personize.ai reframed accuracy as company- and domain-specific and built proactive customer memory.

“while the generated text might be aesthetically pleasing and well-written, it didn't necessarily convey meaningful or accurate information about the customer.”summary, stated

Autonomous batch personalization can become expensive quickly.

Root cause: Multi-agent systems operating over big databases make many tool calls and process large unstructured data volumes.

Resolution: The architecture uses proactive memory and compressed context to reduce repeated context loading.

“multi-agent systems operating on big databases are "super expensive if you don't plan for it and if you don't optimize for it."”summary, stated

Evaluation

Early testing used B2B company feedback and business-specific accuracy judgments rather than generic text-quality metrics.

Metrics: 20+ B2B companies · 5-10 minutes deep research · three-month testing period

“Early testing with 20+ B2B companies showed the system can perform deep research in 5-10 minutes”summary, stated

Outcomes

Cortex generated personalized content for emails, websites, and product pages.
The proactive memory system improved consistency across agents.
Early tests reported senior-level, domain-specific output without human-in-the-loop execution.

The case reports 20+ B2B companies, a three-month test period, and 5-10 minute deep research.

“generate highly personalized, domain-specific content that matches senior-level quality without human-in-the-loop intervention.”summary, stated