Kakao unveiled a personalisation and safety design that distinguishes between the scope of using a user's memory and the authority an AI agent can have to act. It plans to provide needed context within the scope of consent, while having actual tasks carried out within conditions the user has delegated.

Kakao on Oct. 13 introduced the personalisation-based technology 'Project Mosaic', agent safety verification and an operating strategy for its in-house AI models at 'if(kakao)26' held at the Kakao AI Campus in Yongin, Gyeonggi Province.

Kakao Chief Technology Officer Jae-ha Song (송재하) said, "Memory can be a basis for judgement, but it does not become authority to act." He said understanding a user's preferences and securing authority to execute tasks such as ordering and reservations should be handled separately.

Distinguishing purchase facts from preferences... providing memory only to consented services

Project Mosaic is a technology Kakao is preparing to continuously use needed context across KakaoTalk, Kanana Space, Kanvas and multiple agents. It aims to reduce the burden on users of having to explain from scratch items they previously chose and work in progress whenever the service or agent changes.

Kakao is first setting standards for linking and use by checking the purposes for collecting data it holds directly and the scope of consent. It organises information generated in a service into memory and passes only information that is permitted to be provided to the service or agent that requested it.

In this process, it distinguishes observed facts from interpretations and records the source and time as well. Song gave an example assuming a user bought children's shoes as a gift for a nephew. He said if an agent makes a definitive judgement about a user's preferences based only on a purchase record, it can make the error of repeatedly recommending children's shoes afterwards. Information that changes over time, such as delivery status or preferences, must also be checked to see if it is still currently valid.

Song said, "If AI remembers me, it must also remember the fact that I can change."

The scope of providing memory is limited based on consent by service. Memory consented for use only in Service A can be provided to Service A, but even for the same user, it is not provided to Service B if there is no consent for use in Service B.

If a user withdraws consent on source information or the scope of use, memory within that scope will no longer be provided from the next request. It is not a structure that automatically retrieves even memory that another service has already received and stored.

Song said preventing future provision and controlling the use of memory already delivered are separate. He said protecting a user's choices even when using already delivered memory is a task for services and the memory-provision platform to solve together.

If it goes beyond the delegated scope, confirm... verifying action plans and tool calls

Kakao also set out a principle that an agent's execution authority should not expand just because it can use memory. It said conditions such as what product can be bought up to what amount must be made clear, and the scope must be maintained even if work is handed off to another agent.

For example, even if an instruction comes in during an order process to switch an out-of-stock item to another product, the agent should not execute it as-is and must confirm if it goes beyond the scope the user delegated.

Song said, "Proceed within the delegated scope and confirm when it goes beyond," adding, "We cannot leave this boundary only to the model's judgement."

Safety verification is also not limited to the final answer. Kakao's direction is to construct situations such as out-of-stock items or price changes in a test environment and examine the execution process, including what tools the agent called based on what judgement. Even if it answers that it did not place an order, an external task could have progressed if it had called an order API earlier.

It will check whether unintended instructions are mixed into a user's request, memory and current information, and recheck the resulting action plan again before execution. At the service stage, it must also verify account permissions and the scope assigned to the task.

Song stressed, "We must evaluate together whether risky execution was blocked and whether normal requests are properly handled." The aim is to verify that it blocks execution that goes beyond the scope while performing permitted tasks, rather than reducing risk by rejecting all requests.

Device-first processing, role-sharing among models... 대응 to large-scale services

Kakao also unveiled a model operating strategy to provide personalised agents to a large number of users. It said it must secure speed and cost efficiency along with response quality, as the system understands user context and carries out multi-step tasks.

Kakao will first handle requests that can be processed on a smartphone with an on-device model before sending them to a server. The strategy aims to reduce response time and narrow the range in which sensitive information is transmitted externally. It uses server models for tasks requiring up-to-date information, complex reasoning or connections to external services.

'AI Conductor' breaks down a user's goal into step-by-step tasks and selects appropriate models and tools and the order of execution. It assigns models according to request difficulty to reduce unnecessary calls to large models and inference costs.

Kakao Foundation Model head performance leader Byeong-seok Noh (노병석) said, "We provide the best results for all requests, but we do not use the largest model for all requests."