Personalization at scale needs agents to research, infer, plan, and write from a shared understanding of each customer.
2
Cortex UCM turns raw customer data into proactive, standardized memories that agents can reuse, search, filter, and version.
3
Shared memory helps agents stay consistent, but access controls for departments and knowledge domains are still a work in progress.
Summary
Hamed Taheri describes Personize.ai's work on personalization for companies with large customer databases. The system uses multiple agents to research customers, infer missing attributes, plan actions, and generate content for emails, websites, and product pages. Taheri argues that retrieval from raw data is not enough. Different agents can retrieve different fragments, respond differently to prompts, and produce inconsistent customer experiences. Personize.ai's Cortex UCM approach creates proactive memories, such as whether a company is B2B or B2C, then stores them with shared, standardized attributes. Agents can reuse those memories for routing, scoring, and content generation. The approach also stores prompts, examples, and company-specific instructions as reusable memory. Taheri says the system is being tested with more than 20 B2B companies. He is clear that organization and department-level memory controls are not finished, especially where agents must be prevented from accessing another department's knowledge.
Personalization requires a sequence of agent tasks
Taheri says personalization is more than writing attractive copy. Agents need to research a customer, make inferences, reason about the situation, sometimes plan, and then write content. The process must run at scale without a person handling each task. Human supervision remains at a high level, but the workflow cannot depend on a human in every interaction. Multiple agents may be built by different people, so they also need to produce consistent results for the same customer.
Accuracy depends on the company's meaning, not just fluent writing
Personize.ai found that polished text could still be useless to a business. Taheri says a result that looks beautiful does not necessarily mean anything. Accuracy therefore has to be defined for the specific company and its work. The system must also deal with unstructured data, expensive batch processing, tool and API latency, and reliability in an autonomous workflow. Customer understanding is the central problem because every agent needs access to the same underlying view of the customer.
Raw retrieval does not create a reliable customer memory
Taheri compares the approach with retrieval-augmented generation and vector databases. These systems let agents retrieve chunks from a large body of data, but access to raw data does not mean the agent understands the customer. Retrieval can depend on prompt wording, and different agents or repeated runs may receive different fragments. The result can be partial and inconsistent, which creates risk when a company delegates customer communication to an autonomous system.
Function calling and MCP-style connections let agents use applications and databases, but Taheri says they introduce another problem. In a large company, an agent may need to connect to many databases, and tool calls can add data that is irrelevant to the task. That extra material increases the context and does not guarantee that agents share the same customer understanding. Different agents may also call the same function in different ways.
Cortex creates memories proactively and with shared attributes
Personize.ai's Cortex approach uses internal agents to infer and synthesize information instead of only storing raw data. Taheri gives the example of identifying whether a company is B2B, B2C, or a direct consumer, even when the source data does not state that directly. The system then stores attributes with standard names for each customer. Agents can search and filter those attributes, and use them for routing and scoring. The goal is a compact, reusable understanding that every agent can access.
Centralized recall gives agents the same reusable context
Taheri says businesses often build agents that each access customer data in their own way. A centralized recall layer gives every agent a common reference to the larger customer picture. Proactive memorization also compresses raw information, leaving more context available for the task itself. The system stores prompts, best practices, instructions, and examples as user-level memory, so teams do not lose that knowledge when an agent or builder changes. Company and contact information can then be reused across tasks.
The system is being tested for fast business-specific personalization
Personize.ai has been experimenting with Cortex for more than 20 B2B companies. Taheri says implementation can start with one click, and deep research typically takes five to ten minutes. The system can generate website sections and blog posts that match a company's domain and brand voice. He presents this as work in progress. The intended benefit is that agents gain awareness of the business context quickly and can recall the relevant customer information without each team rebuilding the same data access.
Access boundaries for departments are still unfinished
In the question period, Taheri says memory isolation needs more than a customer-level boundary. Personize.ai is working on schema-based structure and expects to add organization and department levels alongside organization-wide and user-specific instructions. The agent can currently examine possible context and choose what it needs, but Taheri says stronger control is required. He describes this part as a work in progress rather than a solved problem.
