[Edaily Reporter Lee So-Hyun ] Just because someone buys children’s shoes online doesn’t necessarily mean they are a parent or particularly interested in children’s shoes. They might have purchased a gift for a niece or nephew. However, if artificial intelligence (AI) judges a user’s preferences based solely on purchase history, it may continue to recommend only children’s shoes in the future. This highlights the difference between AI remembering a user’s behavior and actually understanding the reasons behind it.

Song Jae-ha, Chief Technology Officer (CTO) of Kakao(035720), explained Kakao’s hyper-personalized AI strategy by citing this example during the “13th Edaily Global AI Forum (GAIF 2026)” held on the 12th at the JW Marriott Dongdaemun in Seoul. He emphasized, “It’s true that the AI remembered me, but it didn’t understand me,” adding, “Purchases are facts, but preferences are interpretations.”

What Kakao is focusing on in the Agent AI competition is not merely the accumulation of information. It is the ability to distinguish and understand user behavior and context, and to translate that understanding into actual service execution. On that day, CTO Song presented KakaoTalk-centered Agent AI strategies under the theme “Intelligence in Every Connection.”

“AI That Remembers Me Must Also Recognize Changing Preferences”

Amid the Agent AI competition, Kakao is advancing “Project Mosaic,” which distinguishes and remembers users’ behaviors and preferences. The approach involves separating verified facts—such as purchases and search history—from preferences interpreted by AI, and then adjusting the basis for those judgments or reflecting changing preferences.

For example, even for the same request—“Find me a restaurant for dinner tomorrow”—it recommends a place conducive to conversation when meeting a college classmate, a place where children feel comfortable when dining with family, or a restaurant with vegan options when dining with a vegan friend.

CTO Song explained, “Personalization doesn’t end with simply changing the AI’s tone of voice,” adding, “Recommendations, plans, and the results of tasks must all change accordingly.” He went on to emphasize, “If an AI remembers me, it must also remember that I can change.”

This memory is utilized when the AI actually processes tasks. For example, if a user requests, “Please reorder what I ordered last time,” the AI checks past purchase history, current sales conditions, coupons, and prices, then proceeds with the order after receiving the user’s approval.

“Kanana Space” serves as the interface for this. The company is preparing a feature that allows users to summon the AI by swiping up with two fingers from any screen in KakaoTalk and assign tasks via voice or text. CTO Song noted, “There are now many AIs that can understand speech, but Kakao’s strength lies in having services and partners right at our side that can act on that understanding.”

The company is also advancing its proprietary AI technology. It has developed “Kanana-o,” which understands images, speech, and text, and is currently developing “Kanana Live,” which can speak while listening. The plan is to select AI models based on the complexity of the request to optimize performance, speed, and cost.

“Memory Is Merely the Basis for Judgment”… AI Execution Authority Belongs to the User

Alongside hyper-personalization, another principle emphasized is the separation of AI’s memory from its authority to act. This means that even if an AI knows a user’s preferences or purchase history, it cannot freely carry out every action.

CTO Song stated, “Memory is the basis for judgment, but it does not grant the authority to act,” adding, “Just because an AI knows a user’s past purchases doesn’t mean it can place an order at any price.”

He pointed out that we must also be wary of situations where a product changes due to instructions from an external page or the judgment of another Agent AI while the Agent AI is delegating tasks to other Agents AI. He explained that tasks should proceed within permitted limits, but if they go beyond those limits, the system should pause to verify; for critical actions such as payment, the Agent AI must obtain the user’s explicit approval.

He emphasized, “Now, failure is measured by actions, not words,” noting that damage can occur not only from incorrect AI responses but also from actual errors in orders or reservations.

Kakao is strengthening its execution safety system by verifying not only AI responses but also action plans and calls to external tools in a test environment. It also utilizes “AI Conductor,” which sequentially connects various services such as search, maps, and reservations. The company’s policy is to evaluate Agent AI performance based on whether tasks were actually completed, the time taken, and the cost incurred.

“AI

for All” with 34 Companies… Evolving Toward “One Agent AI per Person”

Kakao is also expanding its Agent AI ecosystem through collaboration with external companies. Thirty-four companies and institutions are participating in “AI for All,” an initiative led by Kakao. Kakao is responsible for user touchpoints, the Agent AI system, and integrated operations, while collaborating with partners on specialized services, AI models, and infrastructure.

“Play MCP,” which connects external functions such as search, inquiry, and booking, currently has approximately 500 tools registered. “Kanvas,” a service that allows individuals to create an agent team on their PC and assign tasks via KakaoTalk, is also in development.

CTO Song stated, “Rather than building every service from scratch, the scope of what we can achieve by connecting with partners who excel in their respective fields is expanding,” adding, “It must also be a viable business for our partners.”

The ultimate goal is “one Agent AI per person,” where each user can configure the Agent AI’s role, memory, and execution permissions. The vision is to enable users to utilize Agent AI tailored to their individual needs, ranging from everyday reservations to work assistance.

CTO Song said, “Users should be able to decide what kind of assistance they receive,” adding, “We’re designing the system so that users can determine—starting from the consent stage—what the AI should remember, which agent to use, and the extent to which they’ll delegate tasks, and be able to change these settings at any time.”

He continued, “There are an increasing number of people who use AI effectively and companies that create excellent AI services,” emphasizing, “I believe this is what enhances our society’s competitiveness. We intend to leverage Kakao’s network to drive this change.”