Multi artificial intelligence agent system

Abstract

The present invention provides to a computing device for constructing a multi artificial intelligence (AI) agent environment, the computing device comprising: a network unit that transmits / receives information in a relationship between at least one user terminal and at least one external resource; a processor that generates at least one clone, generates at least one channel that is a logical space in which the generated at least one clone is logically implemented, and calls the at least one clone to the generated at least one channel; and a memory unit storing information about the at least one clone and the at least one channel.

Description

Multi-AI Agent System

[0001] The present invention relates to an artificial intelligence (AI) agent system. More specifically, the present invention relates to a multi-AI agent environment in which a plurality of AI agents are implemented.

[0002] With the recent rapid advancement of Artificial Intelligence (AI) technology, various AI agent systems have been developed and are being put into practical use. In particular, AI systems based on Large Language Models (LLMs), such as OpenAI's GPT series, Google's Gemini, Anthropic's Claude, and Meta's LLaMA, are demonstrating innovative performance, enabling them to perform various tasks—including natural language understanding and generation, translation, summarization, and code writing—at a level approaching human capabilities.

[0003] In the case of single-agent-based conversational systems, most currently commercialized AI services are based on a structure that generates responses by receiving system and user prompts. These systems can access external Application Programming Interfaces (APIs), such as calendars, document repositories, and web searches, through plugin or tool calling functions. Furthermore, they utilize Retrieval-Augmented Generation (RAG) technology to search for relevant documents using vector embeddings and vector databases, and use this as context to improve response quality.

[0004] In the field of multi-agent systems, active research is underway on multiple AI agents collaborating to solve complex problems. Representative examples include Microsoft’s AutoGen and CrewAI, where agents specialized by role—such as task decomposition, planning, execution, and evaluation agents—interact via centralized management or distributed message passing. Communication between agents takes place through a public bus, conversational channels, or proxy components, and each agent possesses independent access rights to prompts, memory, and tools.

[0005] In terms of personalized AI services, technologies that provide customized services by learning users' preferences or usage patterns are advancing. Virtual assistants such as Siri, Alexa, or Google Assistant implement personalization by maintaining user-specific system prompts, preference memories, and conversation history summaries, while some services offer the ability to pre-define roles, speech patterns, or tool access permissions through custom GPT or character AI.

[0006] Regarding external resource integration technologies, systems are being implemented where AI agents access various external APIs or databases to utilize real-time information. In particular, API integration with AI agents, search engines such as Google Drive, and various AI models is expanding the capabilities of AI agents.

[0007] The objective of the present invention is to provide a multi-AI agent system in which multiple AI agents reflecting the user's personality and knowledge perform various tasks.

[0008] The objective of the present invention is to provide a multi-AI agent environment composed of multiple AI agents capable of active thinking and distributed collaboration.

[0009] To solve the problem of the present invention, an artificial intelligence agent implemented on a server for establishing a multi-artificial intelligence (AI) agent environment comprises: a network unit for transmitting / receiving information in relation to at least one user terminal, the server, and at least one other artificial intelligence agent; a processor unit for generating a response to a user's request received from the user terminal by utilizing a predetermined external resource accessible through the server; and a memory unit for storing information obtained based on interaction with the at least one user terminal or the at least one other artificial intelligence agent—hereinafter referred to as context information—wherein the network unit, the processor unit, and the memory unit are configured to operate by being allocated the computational resources of the server, and the response generated by the processor unit is generated based on the context information.

[0010] According to one embodiment of the present invention, the processor unit may be configured to generate at least one post related to the context information based on the acquired context information, and to post the generated at least one post on a social timeline built on the server.

[0011] According to one embodiment of the present invention, the processor unit may be configured to search for information regarding a certain music as a response to the user request, provide the searched information to the user, and provide a certain connection link that can verify the information.

[0012] According to one embodiment of the present invention, the processor unit may be configured to search for information regarding a predetermined fashion trend as a response to the user request, provide the searched information to the user, and recommend a predetermined outfit to the user based on the information.

[0013] According to one embodiment of the present invention, the processor unit is configured to search for information regarding a predetermined accommodation as a response to the user request and to provide the searched information to the user, and the information may include reservation information regarding the accommodation.

[0014] According to one embodiment of the present invention, the processor unit is configured to search for information about at least one other artificial intelligence agent as a response to the user request, provide the searched information to the user, and perform interaction with the artificial intelligence agent selected by the user among the at least one other artificial intelligence agent, and the information may be subject to homomorphic encryption.

[0015] According to one embodiment of the present invention, the processor unit may be configured to analyze an image or video provided by the user as a response to the user request, generate a predetermined record result related to the image or video based on the analysis result, and provide the generated record result to the user.

[0016] According to one embodiment of the present invention, the processor unit may be configured to calculate a predetermined expected waiting time based on the context information and to provide the response to the user request to the user after waiting for the calculated expected waiting time.

[0017] To solve the problem of the present invention, a method for constructing a multi-artificial intelligence (AI) agent environment is provided, comprising: a step of extracting a predetermined scenario from a target interaction performed between at least one user and at least one AI agent; and a step of applying the extracted scenario to a target interaction performed between at least one user and at least one AI agent.

[0018] According to one embodiment of the present invention, the target interaction may be configured as a sequential combination of at least one sub-interaction.

[0019] According to one embodiment of the present invention, the scenario is configured as a sequential combination of at least one sub-step, and each of the at least one sub-step may correspond to each of the at least one sub-interaction.

[0020] According to one embodiment of the present invention, the target interaction is configured as a sequential combination of at least one sub-interaction, and each of the at least one sub-step constituting the scenario can be sequentially applied to each of the at least one sub-interaction constituting the target interaction.

[0021] To solve the problem of the present invention, a nomadic artificial intelligence (AI) agent system comprises: a local server for generating an AI agent—hereinafter referred to as a clone—and a local device for generating a nomadic AI agent corresponding to the clone—hereinafter referred to as a nomadic AI environment; wherein the nomadic AI environment comprises a nomadic AI execution module configured to receive information related to the clone—hereinafter referred to as context information—from the local server and to generate and execute the nomadic AI agent based on the received context information; and wherein the nomadic AI agent is configured to be driven by being allocated computational resources of the nomadic AI execution module and to generate a response to a user request based on the context information.

[0022] According to one embodiment of the present invention, the nomadic artificial intelligence execution module includes a network unit including an MCP interface; and the nomadic artificial intelligence agent may be configured to access a predetermined operation module included in the nomadic artificial intelligence environment through the MCP interface.

[0023] According to one embodiment of the present invention, the nomadic artificial intelligence agent is configured to generate the response by utilizing a predetermined large-scale language model stored in the nomadic artificial intelligence execution module, and the large-scale language model may be trained based on training data for controlling the operation of the nomadic artificial intelligence environment.

[0024] The multi-artificial intelligence agent system according to the present invention can provide a multi-artificial intelligence agent environment in which multiple artificial intelligence agents reflecting the user's personality and knowledge perform various tasks.

[0025] The multi-artificial intelligence agent system according to the present invention provides a multi-artificial intelligence agent environment composed of a plurality of artificial intelligence agents capable of active thinking and distributed collaboration.

[0026] FIG. 1 is a conceptual diagram showing a multi-artificial intelligence agent system according to the present invention.

[0027] FIG. 2 is a block diagram showing a local server and an external system according to the present invention.

[0028] FIG. 3 is a diagram showing the process of creating a clone and the configuration of the created clone according to the present invention.

[0029] FIG. 4 is a diagram illustrating the process of a local server creating a clone based on a clone creation prompt according to the present invention.

[0030] FIG. 5 is a diagram showing the process of a local server determining the personality of a clone based on biometric information according to the present invention.

[0031] FIG. 6 is a diagram illustrating the process of a local server determining a profile image for a clone based on a profile image generation prompt according to the present invention.

[0032] FIG. 7 is a diagram showing a local server calling a clone in response to a user's request on a channel according to the present invention.

[0033] FIG. 8 is a diagram showing a clone called in response to a user's request on a channel according to the present invention responding.

[0034] FIG. 9 is a diagram showing the interaction between a plurality of users and a plurality of clones on a channel according to the present invention.

[0035] FIG. 10 is a diagram showing that the first clone responds to a user's request based on the task processing of the second clone and the third clone called by a call request on a channel according to the present invention.

[0036] FIG. 11 is a diagram showing different clones sharing data with each other according to the present invention.

[0037] FIG. 12 is a diagram showing a plurality of users sharing information and opinions with one another using clones according to the present invention.

[0038] FIG. 13 is a diagram showing that a clone according to the present invention adjusts a prompt based on a user's request, prompt information, and context information.

[0039] FIG. 14 is a diagram showing that a clone according to the present invention delivers a validated link among the links obtained through an external system to a user.

[0040] FIG. 15 is a diagram showing a clone selected by an orchestrator clone according to the present invention responding to a user's request.

[0041] FIG. 16 is a diagram showing that a clone responds to a user's request based on sub-dialogue information selected by a dialogue manager clone according to the present invention.

[0042] FIG. 17 is a diagram showing that a clone according to the present invention summarizes and transmits the results of interaction with a first user to a second user.

[0043] FIG. 18 is a diagram illustrating the operation of a clone according to the present invention storing context information in a memory unit in the form of triple ordered pairs.

[0044] FIG. 19 is a diagram illustrating the operation of a clone according to the present invention storing context information in a memory unit in the form of a tuple containing triple ordered pairs.

[0045] FIG. 20 is a diagram showing the pattern of how the weights of a tuple according to the present invention change over time based on corresponding initial weights and damping factors.

[0046] FIG. 21 is a diagram showing that the damping factor of a tuple according to the present invention is updated based on repeated input of the same data.

[0047] FIG. 22 is a diagram showing that the weights and damping factors of a tuple according to the present invention are updated based on the reliability of the data.

[0048] FIG. 23 is a diagram showing an embodiment of a social timeline service by a multi-artificial intelligence agent system according to the present invention.

[0049] FIG. 24 is a drawing showing an embodiment of a music recommendation service provided by a clone according to the present invention.

[0050] FIG. 25 is a drawing showing an embodiment of a fashion coordinator service provided by a clone according to the present invention.

[0051] FIG. 26 is a drawing showing an embodiment of an accommodation reservation service provided by a clone according to the present invention.

[0052] FIG. 27 is a drawing illustrating an embodiment of a user matching service provided by a clone according to the present invention.

[0053] FIG. 28 is a drawing illustrating an embodiment of an event recording service provided by a clone according to the present invention.

[0054] FIG. 29 is a diagram illustrating an embodiment of a proxy discussion service provided by a proxy clone according to the present invention.

[0055] FIG. 30 is a diagram showing the operation of determining the firing time of a clone according to the present invention.

[0056] FIG. 31 is a diagram showing that a scenario is extracted from an interaction between at least one user and / or at least one clone according to the present invention.

[0057] FIG. 32 is a diagram showing that a scenario is applied to the interaction between at least one user and / or at least one clone according to the present invention.

[0058] FIG. 33 is a diagram showing a nomadic artificial intelligence agent implemented in a nomadic artificial intelligence environment according to the present invention.

[0059] Hereinafter, a multi-artificial intelligence agent system according to an embodiment of the present invention will be described in detail with reference to the attached drawings. However, it will be readily apparent to those skilled in the art that the attached drawings are provided merely to facilitate the disclosure of the content of the present invention, and that the scope of the present invention is not limited to the scope of the attached drawings.

[0060]

[0061] Multi-AI Agent System

[0062] FIG. 1 is a conceptual diagram showing a multi-artificial intelligence agent system according to the present invention.

[0063] According to one embodiment of the present invention, a multi-artificial intelligence (AI) agent system (10) may refer to any system configured to perform a specific task based on a plurality of artificial intelligence agents configured to interact or cooperate within the same environment or a distributed environment.

[0064] In the specification of the present invention, the artificial intelligence agent may be interchangeably named Clone (2000).

[0065] As illustrated in FIG. 1, the multi-artificial intelligence agent system (10) may include a user terminal (100) and a local server (200).

[0066] Specifically, the local server (200) may be intended to establish a multi-artificial intelligence agent environment. More specifically, the local server (200) may be configured to create at least one channel (1000) and at least one clone (2000), and to manage the created at least one channel (1000) and at least one clone (2000). The channel (1000) may refer to a logical space in which the created clone (2000) is logically implemented. The clone (2000) is a logical entity implemented in the channel (1000) and may be configured to perform a specific task required by a user by utilizing the computational resources of the local server (200), etc.

[0067] Specifically, the user (110) can transmit and receive information in relation to the local server (200) and at least one clone (2000) through the user terminal (100). More specifically, the user (110) can transmit certain information to the local server (200) through the user terminal (100), and the local server (200) can create at least one clone (2000) based on the transmitted information. Additionally, the user (110) can connect to a channel (1000) created by the local server (200) through the user terminal (100) and interact with at least one clone (2000) on the connected channel (1000).

[0068]

[0069] local server

[0070] FIG. 2 is a block diagram showing a local server and an external system according to the present invention.

[0071] As illustrated in FIG. 2, the local server (200) may include a network section (210), a processor section (230), and a memory section (250).

[0072] Specifically, the network unit (210) may be for transmitting / receiving information in relation to at least one user terminal (100) and at least one external resource. Additionally, the network unit (210) may be for transmitting / receiving information in relation to at least one clone (2000) created by the local server (200).

[0073] Specifically, the processor unit (230) may be configured to create at least one clone (2000), create at least one channel (1000) which is a logical space in which the created at least one clone (2000) is logically implemented, and call at least one clone to the created at least one channel.

[0074] Specifically, the memory unit (250) may be configured to store information about at least one clone (2000) and at least one channel (1000). Additionally, the memory unit (250) may include software code for at least one tool (Tool) for performing a task required by a user and software code for at least one artificial intelligence model.

[0075] According to one embodiment of the present invention, a local server (200) may call an Application Programming Interface (API) for at least one tool or at least one artificial intelligence model for performing a task requested by a user from a predetermined external resource. Additionally, the local server (200) may obtain information used to perform a task requested by a user from a predetermined external resource.

[0076] According to one embodiment of the present invention, the external resource may include at least one resource selected from the group comprising a storage service resource, a social network service resource, a search engine service resource, and an artificial intelligence model resource.

[0077] For example, a storage service resource may refer to at least one resource selected from a group including Google Drive, Dropbox, Microsoft OneDrive, and Amazon S3 (Amazon Simple Storage Service), etc. However, it is not limited thereto.

[0078] For example, a social network service resource may mean at least one resource selected from the group including Facebook, X (formerly Twitter), Instagram, and LinkedIn, etc. However, it is not limited thereto.

[0079] For example, a search engine service resource may mean at least one resource selected from a group including Google, Naver, Bing, and Yahoo, etc. However, it is not limited thereto.

[0080] For example, an AI model resource may refer to at least one resource selected from a group including OpenAI's GPT, Google's Gemini, Anthropic's Claude, Meta's LLaMA, and Stability AI's Stable Diffusion, etc. However, it is not limited thereto.

[0081]

[0082] clone

[0083] FIG. 3 is a diagram showing the process of creating a clone and the configuration of the created clone according to the present invention.

[0084] As illustrated in FIG. 3, the clone (2000) can be created by a local server (200).

[0085] Specifically, the local server (200) may receive at least one piece of information selected from a group including a clone creation prompt and biological information of the user (110) from at least one user terminal (100). The local server (200) may create a clone (2000) based on at least one piece of information received.

[0086] As illustrated in FIG. 3, the clone (2000) may include a network unit (2100), a processor unit (2300), and a memory unit (2500). Here, the network unit (2100), the processor unit (2300), and the memory unit (2500) may each be configured to operate by being allocated at least some of the resources of the network unit (210), the processor unit (230), and the memory unit (250) of the local server (200).

[0087] Specifically, the network unit (2100) may be for transmitting / receiving information in relation to at least one user terminal (100), a local server (200), and at least one other clone (2000).

[0088] Specifically, the processor unit (2300) may be configured to generate a response to a user request received from a user terminal (100). More specifically, the processor unit (2100) may utilize software code for at least one tool and software code for at least one artificial intelligence model stored in the memory unit (250) of the local server (200) in the process of generating the response. Additionally, the processor unit (2100) may utilize an API for at least one tool or at least one artificial intelligence model called by the local server (200) from an external resource and information obtained from an external resource in the process of generating the response.

[0089] Specifically, the memory unit (2500) may be configured to store information collected from at least one user terminal (100) with which the clone (2000) interacted and at least one channel (1000) in which the clone (2000) is implemented. In the specification of the present invention, the information collected from at least one user terminal (100) with which the clone (2000) interacted and at least one channel (1000) in which the clone (2000) is implemented may be interchangeably referred to as context information.

[0090]

[0091] clone creation

[0092] Clone creation prompt

[0093] FIG. 4 is a diagram illustrating the process of a local server creating a clone based on a clone creation prompt according to the present invention.

[0094] According to one embodiment of the present invention, the memory unit (250) of the local server (200) may include a prompt storage unit (251). Specifically, the prompt storage unit (251) may store at least one prompt for creating a clone (2000) (e.g., see (B) in FIG. 4).

[0095] As illustrated in FIG. 4, the local server (200) may be configured to receive a clone creation prompt (e.g., see (A) in FIG. 4) from the user terminal (100) and to create a clone (2000) based on the received clone creation prompt.

[0096] Specifically, the processor unit (230) of the local server (200) can output a final prompt (e.g., see (C) in FIG. 4) by using a clone creation prompt and at least one prompt stored in the prompt storage unit (251) as inputs to a predetermined Large Language Model (LLM). Additionally, the processor unit (230) can generate a clone (2000) based on the output final prompt.

[0097] As illustrated in FIG. 4, the output final prompt can be stored in the prompt storage unit (251). Specifically, the stored final prompt can be input into a large language model along with another clone creation prompt received from the user terminal (100) during the clone (2000) creation process described above, and used to output another final prompt.

[0098] According to one embodiment of the present invention, a user (110) can generate a desired clone (2000) by transmitting a clone generation prompt containing a description of the desired clone (2000) to a local server (200) through a user terminal (100). In the process of generating the clone (2000), the local server (200) can generate a final prompt by improving the clone generation prompt transmitted by the user (110) using at least one prompt previously stored in a prompt storage unit (251).

[0099]

[0100] Biometric information

[0101] FIG. 5 is a diagram showing the process of a local server determining the personality of a clone based on biometric information according to the present invention.

[0102] According to one embodiment of the present invention, biological information regarding a user (110) may include biometric information and medical information. However, it is not limited thereto.

[0103] As illustrated in FIG. 5, the local server (200) may be configured to receive biometric information about the user (110) from the user terminal (100) and to create a clone (2000) based on the received biometric information.

[0104] Specifically, the processor unit (230) of the local server (200) can extract at least one biological feature by using it as input to a feature extraction model of biometric information. Additionally, the processor unit (230) can extract a persona from the extracted biological feature through biological feature-persona matching using a persona extraction model. Additionally, the processor unit (230) can output a clone creation prompt by using the extracted persona as input to a predetermined large-scale language model. Additionally, the processor unit (230) can create a clone (2000) based on the output clone creation prompt.

[0105] Here, the clone creation prompt may mean a prompt for determining the personality of the clone (2000) being created. That is, the user (110) can create a clone (2000) having a persona or personality corresponding to their biometric information by utilizing their biometric information, etc.

[0106]

[0107] Profile Image

[0108] FIG. 6 is a diagram illustrating the process of a local server determining a profile image for a clone based on a profile image generation prompt according to the present invention.

[0109] According to one embodiment of the present invention, a profile image may refer to an identification image displayed together with or in place of a clone (2000) on an interface unit (not shown), such as a user terminal (100).

[0110] According to one embodiment of the present invention, the memory unit (250) of the local server (200) may further include an image storage unit (253). Specifically, the image storage unit (253) may store information regarding a pair of at least one profile image and a corresponding image space vector.

[0111] As illustrated in FIG. 6, the local server (200) may be configured to receive a profile image generation prompt (e.g., see (A) in FIG. 6) from the user terminal (100) and to generate a profile image of the clone (2000) based on the received profile image generation prompt.

[0112] Specifically, the processor unit (230) of the local server (200) can output a prompt space vector corresponding to a profile image by using it as input to a vector embedding model of a profile image generation prompt. Additionally, the processor unit (230) can select at least one profile image for a clone (2000) from the image storage unit (253) through matching between the output prompt space vector and the ordered pair stored in the image storage unit (253).

[0113] According to one embodiment of the present invention, the profile image included in the ordered pair stored in the image storage unit (253) may be an image generated by the processor unit (230) using a profile image generation prompt received from the user terminal (100) as input to a predetermined text-image artificial intelligence model. Additionally, the image space vector included in the ordered pair stored in the image storage unit (253) may be a vector output by the processor unit (230) using the aforementioned profile image generation prompt as input to a vector embedding model. That is, the profile image included in the ordered pair and the corresponding space vector may originate from the same profile image generation prompt.

[0114] According to one embodiment of the present invention, a user (110) can set a desired profile image by transmitting a profile image creation prompt containing a description of the profile image of a desired clone (2000) to a local server (200) through a user terminal (100). In the profile image setting process, the local server (200) can select a suitable profile image corresponding to the profile image creation prompt without involving the creation of a new profile image by using at least one sequence pair previously stored in an image storage unit (253).

[0115]

[0116] QR code-based clone generation

[0117] According to one embodiment of the present invention, a clone (2000) may be generated based on a predetermined Automatic Identification and Data Capture (AIDC) means. Specifically, the automatic identification and data capture means may be used as an identification token to identify or call a clone generation prompt for generating the clone (2000), information about the persona or personality of the clone (2000), or a profile image generation prompt for determining the profile image of the clone (2000), as initial setting parameters for generating the clone (2000).

[0118] According to one embodiment of the present invention, the automatic identification and data collection means may include an optical-based recognition code and a non-optical-based recognition means. Specifically, a user (110) may generate a predetermined clone (2000) corresponding to the optical-based recognition code or the non-optical-based recognition means by capturing the optical-based recognition code through a user terminal (100) or by contacting the user terminal (100) with the non-optical-based recognition means.

[0119] For example, optical-based identification codes may include linear barcodes including UPC (Universal Product Code), EAN (European Article Number), Code 39, Code 128, ITF (Interleaved Two of Five), Codabar, MSI Plessey, RSS-14 / GS1 DataBar, or Pharmacode (barcode for pharmaceutical packaging), and matrix codes including QR Code (Quick Response Code), Data Matrix, Aztec Code, PDF417, MaxiCode, Micro QR Code, Han Xin Code (Chinese Standard 2D Code), or VeriCode / VSCode (Industrial Identification Code). However, they are not limited thereto.

[0120] For example, non-optical based recognition means may include RFID (Radio Frequency Identification), NFC (Near Field Communication), BLE beacons (Bluetooth Low Energy Beacon), or magnetic stripes. However, they are not limited thereto.

[0121]

[0122] Regarding clone operation

[0123] Clone management and calling

[0124] FIG. 7 is a diagram showing a local server calling a clone in response to a user's request on a channel according to the present invention.

[0125] According to one embodiment of the present invention, a user request may mean that a user (110) requests the performance of a specific task on a predetermined channel (1000). Additionally, a user request may mean that a user (110) requests the call of a specific clone (2000) on the channel (1000). Additionally, a user request may mean that a user (110) connects to a specific channel (1000) where a specific clone (2000) is exclusively implemented, and in this case, the connection of the user (110) may be considered as a request for the call of the specific clone (2000). However, it is not limited thereto.

[0126] As illustrated in FIG. 7, when the above-described user request is input on the channel (1000), the local server (200) may call at least one clone (2000) corresponding to the user request on the channel (1000). Specifically, calling the clone (2000) may mean logically implementing the clone (2000), which is a logical entity, on the channel (1000). More specifically, the processor unit (230) of the local server (200) may logically implement the corresponding clone (2000) on the channel (1000) based on information about at least one clone (2000) stored in the memory unit (250). Here, the information about the clone (2000) may mean software code about the clone (2000) and context information collected in the state where the clone (2000) is implemented. That is, when the local server (200) executes the software code described above by combining the context information described above, a clone (2000) corresponding to the software code and context information may be called.

[0127]

[0128] Clone's response to user request

[0129] FIG. 8 is a diagram showing a clone called in response to a user's request on a channel according to the present invention responding.

[0130] As illustrated in FIG. 8, at least one clone (2000) called on the channel (1000) can generate a predetermined response in response to a user request (see (A) in FIG. 8).

[0131] Specifically, the clone (2000) can generate a response by executing software code for at least one tool and software code for at least one artificial intelligence model stored in the memory section (250) of the local server (200). For example, the clone (2000) can generate a response based on a predetermined large-scale language model, but is not limited thereto.

[0132] Specifically, the clone (2000) can generate a response using context information collected from at least one user terminal (1000) and / or at least one channel (1000).

[0133] Specifically, the clone (2000) can access a specific external resource directly or via a local server (200) and generate a response by utilizing the specific external resource.

[0134] According to one embodiment of the present invention, the external resource may include at least one resource selected from the group comprising a storage service resource, a social network service resource, a search engine service resource, and an artificial intelligence model resource.

[0135] For example, a storage service resource may refer to at least one resource selected from a group including Google Drive, Dropbox, Microsoft OneDrive, and Amazon S3 (Amazon Simple Storage Service), etc. However, it is not limited thereto.

[0136] For example, a social network service resource may mean at least one resource selected from the group including Facebook, X (formerly Twitter), Instagram, and LinkedIn, etc. However, it is not limited thereto.

[0137] For example, a search engine service resource may mean at least one resource selected from a group including Google, Naver, Bing, and Yahoo, etc. However, it is not limited thereto.

[0138] For example, an AI model resource may refer to at least one resource selected from a group including OpenAI's GPT, Google's Gemini, Anthropic's Claude, Meta's LLaMA, and Stability AI's Stable Diffusion, etc. However, it is not limited thereto.

[0139]

[0140] Interaction between multiple users and multiple clones

[0141] FIG. 9 is a diagram showing the interaction between a plurality of users and a plurality of clones on a channel according to the present invention.

[0142] As illustrated in FIG. 9, multiple users (110-1 to 110-N) can be connected to the same channel (1000). Likewise, multiple clones (2000-1 to 2000-M) can be called on the same channel (1000).

[0143] Specifically, multiple users (110-1 to 110-N) and multiple clones (2000-1 to 2000-M) can interact with each other on the channel (1000).

[0144] For example, multiple users (110-1 to 110-N) can interact with each other on the channel (1000).

[0145] For example, multiple clones (2000-1 to 2000-M) can interact with each other on the channel (1000).

[0146] For example, a single clone (2000) can generate a response for a request from a single user (110).

[0147] For example, for a request from a single user (110), at least some of the multiple clones (2000-1 to 2000-M) may generate a response.

[0148] For example, a single clone (2000) can generate a response to at least some of the requests of multiple users (110-1 to 110-N).

[0149] For example, at least some of the multiple clones (2000-1 to 2000-M) may generate a response to a request from at least some of the multiple users (110-1 to 110-N).

[0150] However, the examples described above are for convenience of explanation only, and the present invention is not limited thereto.

[0151]

[0152] Distributed collaboration between clones

[0153] FIG. 10 is a diagram showing that the first clone responds to a user's request based on the task processing of the second clone and the third clone called by a call request on a channel according to the present invention.

[0154] As illustrated in FIG. 10, a clone (2000) may utilize the task processing results of at least one other clone (2000) in the process of generating a response to a user request. In the specification of the present invention, the above-described response generation process may be referred to as decentralized collaboration between clones (2000).

[0155] Specifically, when a first clone (2000-1) called first on a channel (1000) receives a predetermined user request from a user (110), the first clone (2000-1) can determine whether it can generate a suitable response to the received user request based on context information it possesses and accessible external resources.

[0156] Specifically, if the first clone (2000-1) determines that it cannot generate a suitable response on its own, the first clone (2000-1) may send a call request to the local server (200) to call at least one other clone (2000) on the channel (1000).

[0157] Specifically, the local server (200) may call at least one clone (2000) on the channel (1000) that possesses context information suitable for generating a response based on a received call request and access rights to external resources. For example, as illustrated in FIG. 10, the local server (200) may call a second clone (2000-2) and a third clone (2000-3).

[0158] Specifically, the first clone (2000-1) may distribute the task of generating a response to a user request to each of the called second clone (2000-2) and third clone (2000-3). The second clone (2000-2) and the third clone (2000-3) may process the respective tasks assigned by the first clone (2000-1) based on the context information and accessible external resources they each possess. The second clone (2000-2) and the third clone (2000-3) may transmit the results of their respective task processing to the first clone (2000-1).

[0159] Specifically, the first clone (2000-1) can generate a response to a user request based on the task processing results received from the second clone (2000-2) and the third clone (2000-3).

[0160] Meanwhile, FIG. 10 illustrates that the first clone (2000-1) calls only two clones (2000-2, 2000-3), but the present invention is not limited thereto.

[0161]

[0162] Data sharing between clones

[0163] FIG. 11 is a diagram showing different clones sharing data with each other according to the present invention.

[0164] According to one embodiment of the present invention, a clone (2000) may be configured to store certain context information through a memory unit (2500). As described above, the context information may refer to all information collected by the clone (2000) from at least one user terminal (100) and at least one channel (2000).

[0165] According to one embodiment of the present invention, a clone (2000) can transmit / receive information in relation to at least one other clone (2000) on or outside the channel (1000) through a network unit (2100).

[0166] As illustrated in FIG. 11, different clones (2000) can exchange at least some of the context information stored in their respective memory units (2500) with one another. Specifically, different clones (2000) can exchange information about at least one user (110) or information about at least one channel (1000) with one another.

[0167] For example, different clones (2000) created by different users (110) may each possess different context information. The context information possessed by the different clones (2000) may include information about each of the different users (110). That is, the different clones (2000) may exchange information about the different users (110). However, they are not limited thereto.

[0168]

[0169] Memory throw function

[0170] FIG. 12 is a diagram showing a plurality of users sharing information and opinions with one another using clones according to the present invention.

[0171] As illustrated in FIG. 12, different users (first user (110-1), second user (110-2)) can share information and opinions with each other through a single clone (2000).

[0172] Meanwhile, FIG. 12 illustrates two users (110-1, 110-2) sharing information and opinions through a single clone (2000), but this is for convenience of explanation only and the present invention is not limited thereto.

[0173] Specifically, the first user (110-1) and the clone (2000) can interact on the first channel (1000-1). The first user (110-1) can input certain information on the first channel (1000-1) (see ① in FIG. 12). The first user (110-1) can request the second user (110-2) to share the input information (see ② in FIG. 12).

[0174] Specifically, the clone (2000) can structure the information entered by the first user (110-1) according to a predetermined method in response to the information sharing request of the first user (110-1) (see ③ in FIG. 12).

[0175] Specifically, the clone (2000) can be called to the second channel (1000-2) connected to the second user (110-2). Here, the second channel (1000-2) is different from the first channel (1000-1), but is not limited thereto. The called clone (2000) can input the result of structuring the information input by the first user (110-1) on the first channel (1000-1) onto the second channel (1000-2) and share it with the second user (110-2) (see ④ in FIG. 12).

[0176] Specifically, the second user (110-2) can write a predetermined response opinion regarding the shared information and input it on the second channel (1000-2) (see ⑤ in FIG. 12).

[0177] Specifically, the clone (2000) can be called on the first channel (1000-1), input a response opinion entered by the second user (110-2) on the second channel (1000-2) onto the first channel (1000-1), and transmit it to the first user (110-1) (see ⑥ in FIG. 12).

[0178] Meanwhile, in the process described above (see ① to ⑥ of FIG. 12), the clone (2000) may be optionally implemented on the first channel (1000-1) or the second channel (1000-2). Alternatively, the clone (2000) may be implemented simultaneously on the first channel (1000-1) and the second channel (1000-2). However, it is not limited thereto.

[0179]

[0180] Adaptive prompting

[0181] FIG. 13 is a diagram showing that a clone according to the present invention adjusts a prompt based on a user's request, prompt information, and context information.

[0182] According to one embodiment of the present invention, the memory unit (2500) of the clone (2000) may store prompt information and context information. Specifically, the prompt information may include format information for converting a user request into a form of input suitable for a predetermined large-scale language model. Details regarding the context information are as described above.

[0183] As illustrated in FIG. 13, the clone (2000) receives a specific user request from a user (110) on the channel (1000) and can generate a response by self-adjusting a prompt corresponding to the received user request.

[0184] Specifically, the processor unit (2300) of the clone (2000) can output an adjustment prompt by using a user request received from the user (110), prompt information and context information stored in the memory unit (2500) as input to a first large-scale language model (not shown). More specifically, the adjustment prompt may mean a prompt that is adjusted into a form suitable for generating a response to a user request by taking into account not only the user request and prompt information stored in the memory unit (2500) but also the context information.

[0185] Specifically, the processor unit (2300) can generate a response to a user request by using the output adjustment prompt as input to a second large-scale language model (not shown).

[0186] According to one embodiment of the present invention, the first large-scale language model (not shown) and the second large-scale language model (not shown) described above may be different models from each other. Alternatively, the first large-scale language model (not shown) and the second large-scale language model (not shown) may be the same model from each other. Alternatively, the processor unit (2300) may utilize a single large-scale language model as the first large-scale language model (not shown) and the second large-scale language model (not shown). However, it is not limited thereto.

[0187] According to one embodiment of the present invention, the adjustment prompt output by the processor unit (2300) can be stored as prompt information in the memory unit (2500). That is, the clone (2000) can be configured to continuously update the prompt information stored in the memory unit (2500) based on context information.

[0188]

[0189] Link validation

[0190] FIG. 14 is a diagram showing that a clone according to the present invention delivers a validated link among the links obtained through an external system to a user.

[0191] As illustrated in FIG. 14, the clone (2000) may utilize certain external resources in the process of generating a response to a user request on the channel (1000). Specifically, the user (110) may request information regarding certain links that lead to a specific website and a specific file. At this time, the clone (2000) may obtain information regarding the link desired by the user (110) from the external resources and generate a response that transmits the information regarding the obtained link to the user (110).

[0192] Meanwhile, the details regarding external resources are as described above.

[0193] As illustrated in FIG. 14, the processor unit (2300) of the clone (2000) can perform validation on information about a link obtained from an external resource.

[0194] Specifically, validation may mean checking whether an acquired link is a valid link that leads to an actually existing site or file, etc. Alternatively, validation may mean checking whether an acquired link is a link that leads to an appropriate site or file, etc. that meets the user request. However, it is not limited thereto.

[0195] Specifically, the processor unit (2300) can perform link validation on links obtained from external resources. Information regarding links that have passed link validation can be transmitted to the user (110) in the form of a response to a user request. That is, the processor unit (2300) can generate a response to a user request based on links whose validity has been verified.

[0196]

[0197] Multi-AI Agent Orchestration

[0198] FIG. 15 is a diagram showing a clone selected by an orchestrator clone according to the present invention responding to a user's request.

[0199] As illustrated in FIG. 15, an orchestrator clone (3000) may be logically implemented on at least one channel (1000).

[0200] According to one embodiment of the present invention, an orchestrator clone (3000) may be configured to select a clone (2000) to generate a response to a user request among a plurality of clones (2000) implemented on a channel (1000). The clone (2000) selection process described above in the specification of the present invention may be referred to as orchestration for a multi-artificial intelligence agent.

[0201] According to one embodiment of the present invention, an orchestrator clone (3000) may use a lightweight language model for performing clone (2000) selection. Specifically, the lightweight language model is a lightweight form compared to the large-scale language model used by a conventional clone (2000), and may be a model solely for clone (2000) selection rather than for generating a response to a user request.

[0202] Specifically, when a user (110) inputs a specific user request on a channel (1000) (see ① in FIG. 15), the orchestrator clone (3000) can select a suitable clone (2000) capable of generating a response to the user request from among a plurality of clones (2000-1, 2000-2, 2000-3). At this time, when selecting the suitable clone (2000), the orchestrator clone (3000) may consider the context information possessed by each clone (2000) and accessible external resources. For example, the orchestrator clone (3000) may select clone (2000-2) as the suitable clone (2000).

[0203] Specifically, the orchestrator clone (3000) can transmit a response instruction to the selected clone (2000-2) (see ② in FIG. 15). The clone (2000-2) that receives the response instruction can generate a response to the user request and input it onto the channel (1000) (see ③ in FIG. 15).

[0204]

[0205] Conversation Manager Clone

[0206] FIG. 16 is a diagram showing that a clone responds to a user's request based on sub-dialogue information selected by a dialogue manager clone according to the present invention.

[0207] As illustrated in FIG. 16, a conversation manager clone (4000) may be logically implemented on at least one channel (1000).

[0208] According to one embodiment of the present invention, the dialogue manager clone (4000) may be for selecting at least one sub-dialogue information (252) that forms the basis of the process in which the clone (2000) generates a response to a user request on the channel (1000).

[0209] According to one embodiment of the present invention, sub-conversation information (252) may be stored in a memory unit (250) of a local server (200). Specifically, the sub-conversation information (252) may be intended to guide the direction and goal of interaction or conversation between a user (110) and a clone (2000) on a channel (1000). More specifically, the sub-conversation information (252) may refer to information regarding a response format or conversation format that serves as a basis for generating a response to a user request.

[0210] Specifically, when a user (110) inputs a predetermined user request on a channel (1000) (see ① in FIG. 16), the conversation manager clone (4000) can select at least one suitable sub-conversation information (252-1 to 252-N) from among a plurality of sub-conversation information (252-1 to 252-N) stored in the memory unit (250) of the local server (200) that can serve as a basis for generating a response to the input user request (see ② in FIG. 16).

[0211] Specifically, the clone (2000) can access the memory section (250) of the local server (200) to obtain sub-dialogue information (252) selected by the dialogue manager clone (4000) (see ③ in FIG. 16).

[0212] Specifically, the clone (2000) can generate a response to a user request based on acquired sub-conversation information (252) and input the generated response onto the channel (1000) (see ④ in FIG. 16). When the clone (2000) inputs the response, the clone (2000) can also provide the sub-conversation information (252) selection history (see (A) in FIG. 16).

[0213]

[0214] Automatic summarization and delivery features

[0215] FIG. 17 is a diagram showing that a clone according to the present invention summarizes and transmits the results of interaction with a first user to a second user.

[0216] As illustrated in FIG. 17, the clone (2000) can automatically summarize the results of a user request or interaction with the user (110) and deliver the summary results to another user (110) who is evaluated to be able to provide a suitable answer to the summary results. The user (110) who receives the summary results can interact with the user (110) who is determined to need an answer through the local server (200) or through other methods.

[0217] For example, when a first user (110-1) interacts with a clone (2000) on a first channel (1000-1) (see ① in FIG. 17), the clone (2000) may evaluate that a second user (110-2) can provide an appropriate response to the first user (110-1)'s request. Accordingly, the clone (2000) may summarize and transmit the first user (110-1)'s request and the results of the interaction with the first user (110-1) to the second user (110-2) on a second channel (1000-2). Upon receiving the summary results, the second user (110-2) may contact the first user (110-1) to provide an appropriate response (see ③ in FIG. 17).

[0218]

[0219] Management of contextual information based on triples

[0220] Basic structure - Triple

[0221] FIG. 18 is a diagram illustrating the operation of a clone according to the present invention storing context information in a memory unit in the form of triple ordered pairs.

[0222] As described above, the clone (2000) may be configured to store certain context information through the memory unit (2500). Specifically, the context information may refer to all information collected by the clone (2000) from at least one user terminal (100) and at least one channel (2000). More specifically, the clone (2000) may collect context information by interacting with at least one user (110) or at least one other clone (2000) while logically implemented on the channel (2000) by the local server (200).

[0223] As illustrated in FIG. 18, the processor unit (2300) of the clone (2000) may be configured to generate at least one triple ordered pair (Triple) based on the content of the clone (2000) interacting with at least one user (110) or at least one other clone (2000) on the channel (1000). Specifically, the processor unit (2300) may generate at least one triple ordered pair by using the content of the interaction as input to a predetermined large-scale language model.

[0224] Here, a triple ordered pair refers to an ordered pair consisting of a subject, a predicate, and an object, and can mean a data structure used to express the main content of an interaction in a condensed manner. For example, a triple ordered pair may be in the form (Tom, like, kpop), but is not limited thereto.

[0225] As illustrated in FIG. 18, the clone (2000) may be configured to store context information collected as a result of interaction in the memory unit (2500) in the form of the aforementioned triple ordered pairs. Specifically, the clone (2000) may store triple ordered pairs generated by the processor unit (2300) in the memory unit (2500).

[0226] As illustrated in FIG. 18, the clone (2000) may be configured to use context information stored in the memory unit (2500) in the form of triple ordered pairs during the process of generating a response to a predetermined request received from the user (110).

[0227] Hereinafter, triple ordered pairs will be described in detail with reference to A, B, C, and D of FIG. 18.

[0228]

[0229] Fig. 18A illustrates an example of the content of an interaction between a clone (2000) and a user (110) or another clone (2000). Specifically, the content of the interaction includes “I like this kpop” and “This favorite band is BTS”.

[0230] Fig. 18B illustrates an example of a triple ordered pair generated by a clone (2000) based on interaction content. Specifically, the clone (2000) can generate two triple ordered pairs using the interaction content as input to a large-scale language model. For example, the triple ordered pairs may be in the form of (Tom, like, kpop) and (Tom favorite band, BTS), etc. The generated triple ordered pairs may be stored in a memory unit (2500).

[0231] Fig. 18 C and D represent a response generated by the clone (2000) in response to a request from the user (110) based on the request from the user (110) and context information stored in the memory unit (2500) in the form of a triple ordered pair. For example, the user (110) may send a request to the clone (2000) in the form of “you play music for me?” (see Fig. 18 C). In response, the clone (2000) may generate a response such as “I'll play your favorite band, BTS music.” using context information stored in the form of a triple ordered pair such as (Tom, like, kpop) and (Tom favorite band, BTS), and provide the generated response to the user (110).

[0232]

[0233] Tuple - Weight and Damping Factor

[0234] FIG. 19 is a diagram illustrating the operation of a clone according to the present invention storing context information in a memory unit in the form of a tuple containing triple ordered pairs.

[0235] As described above, the clone (2000) can generate a predetermined triple sequence pair based on interaction with at least one user (110) or at least one other clone (2000).

[0236] As illustrated in FIG. 19, the clone (2000) may be configured to generate a tuple based on the generated triple ordered pair. Specifically, the tuple may be a combined representation between the triple ordered pair and metadata including a weight and a decay factor (Df) for the triple ordered pair. Here, the weight is a factor that determines the importance of contextual information corresponding to the triple ordered pair, and the higher the weight of the triple ordered pair, the higher the importance of the corresponding contextual information. Additionally, the decay factor is a factor that indicates the rate at which the weight of the triple ordered pair decreases over time, and the higher the decay factor of the triple ordered pair, the faster the importance of the corresponding contextual information decreases. The clone (2000) may determine the weight and decay factor corresponding to the generated triple ordered pair and generate a tuple corresponding to the triple ordered pair by combining the triple ordered pair with the determined weight and decay factor. The generated tuple can be stored in the memory unit (2500).

[0237] FIG. 20 is a diagram showing the pattern of how the weights of a tuple according to the present invention change over time based on corresponding initial weights and damping factors.

[0238] According to one embodiment of the present invention, the weight may decrease over time. Specifically, the weight may be expressed by the following [Equation 1].

[0239] [Mathematical Formula 1]

[0240]

[0241] (Here,

[0242] W T : Weights for triple ordered pairs at time T,

[0243] W T=0 : Initial weight of the triple ordered pair at time T0 (the time when the triple ordered pair was first generated),

[0244] T: Time elapsed since the time the triple ordered pair was first created,

[0245] Df: Damping factor for triple ordered pairs)

[0246]

[0247] As illustrated in FIG. 20, the tuples generated by the clone (2000) may be assigned different initial weights and damping factors at the time of initial generation. For example, the initial weight of the first tuple may be greater than the initial weight of the second tuple. Additionally, the damping factor of the first tuple may be greater than the damping factor of the second tuple. Here, because the damping factor of the first tuple is greater than the damping factor of the second tuple, the weight of the first tuple may decrease at a higher rate compared to the weight of the second tuple. Thus, even though the initial weight of the first tuple was greater than the initial weight of the second tuple, after a certain time has passed, the weight of the first tuple may be smaller than the weight of the second tuple.

[0248] According to one embodiment of the present invention, a clone (2000) can generate a response to a user (110) using context information stored in the memory unit (2500) in the form of a tuple. Specifically, the clone (2000) may be configured to generate a response by processing a tuple with a high weight as more important information compared to a tuple with a low weight.

[0249]

[0250] Update of weights and damping factors

[0251] Update of damping factor based on repeated input

[0252] FIG. 21 is a diagram showing that the damping factor of a tuple according to the present invention is updated based on repeated input of the same data.

[0253] According to one embodiment of the present invention, if the content of an interaction performed at different times is equivalent, the damping factor of a tuple that has been generated and stored in the memory unit (2500) can be updated. Specifically, when a predetermined tuple is generated as a result of an interaction performed at a predetermined time and stored in the memory unit (2500), and an identical triple ordered pair corresponding to the tuple is generated as a result of an interaction performed at any time after that time, the processor unit (2300) can update the damping factor assigned to the generated tuple. More specifically, the processor unit (2300) can reduce the damping factor assigned to the generated tuple. Accordingly, in response to the repeated generation of identical triple ordered pairs as a result of an interaction, the clone (2000) can reduce the rate or rate at which the weight of the tuple corresponding to the repeatedly generated triple ordered pair decreases over time.

[0254] Hereinafter, the update of the damping factor will be explained in detail with reference to FIG. 21.

[0255]

[0256] As illustrated in FIG. 21, at time T1, the clone (2000) may receive information “is my favorite band” from the user (110) or the clone (2000) through a predetermined interaction. The clone (2000) may generate a triple ordered pair (Tom, favorite band, BTS) (A) based on the received information. Subsequently, the clone (2000) may generate a tuple (B) based on the generated triple ordered pair and store the generated tuple (B) in the memory unit (2500). At this time, the weight and damping factor assigned to the tuple (B) may be 10 and 0.7, respectively.

[0257] As illustrated in FIG. 21, at time T2, the clone (2000) may receive information “BTS is my favorite band” from the user (110) or the clone (2000) through a predetermined interaction. The clone (2000) may generate a triple ordered pair (Tom, favorite band, BTS) (C) based on the received information. At this time, the generated triple ordered pair (C) may be composed of the same subject, predicate, and object as the triple ordered pair (A) or tuple (B) generated at time T1.

[0258] As illustrated in FIG. 21, in response to the generation of identical triple ordered pairs (A, C) at time points T1 and T2, the clone (2000) can generate an updated tuple (D) by updating the decay factor of the tuple (B) stored in the memory unit (2500). Specifically, the decay factor of the updated tuple (D) can be changed from the existing 0.7 to 0.63, which is 10% lower. The updated updated tuple (D) can be stored in the memory unit (2500).

[0259]

[0260] Update of weights and damping factors based on information source

[0261] FIG. 22 is a diagram showing that the weights and damping factors of a tuple according to the present invention are updated based on the reliability of the data.

[0262] According to one embodiment of the present invention, the weight of a tuple that has been generated and stored in the memory unit (2500) can be updated based on the reliability of the data provided to the clone (2000) through interaction. Specifically, when a predetermined tuple is generated as a result of an interaction at a predetermined time and stored in the memory unit (2500), and an identical triple ordered pair corresponding to the tuple is generated as a result of an interaction at some point after that time, and the reliability of the data obtained through the interaction is determined to be greater than or equal to a predetermined threshold, the processor unit (2300) can update the weight and attenuation factor assigned to the generated tuple. More specifically, the processor unit (2300) can increase the weight assigned to the generated tuple and decrease the assigned attenuation factor. Accordingly, in response to the determination that the reliability of the data obtained through the interaction is above a predetermined threshold value along with the repeated generation of the same triple ordered pair as a result of the interaction, the clone (2000) can reduce the rate or rate at which the weight of the tuple corresponding to the repeatedly generated triple ordered pair decreases over time, and at the same time, the weight of the corresponding tuple itself can be adjusted upward.

[0263] For example, data reliability may refer to the source of the data, but is not limited to this.

[0264] Below, the updating of weights and damping factors will be explained in detail with reference to FIG. 22.

[0265]

[0266] As illustrated in FIG. 22, at time T1, the clone (2000) may receive information “likes kpop music” from the user (110) through a predetermined interaction. The clone (2000) may generate a triple ordered pair (John, like, kpop) (A) based on the received information. Subsequently, the clone (2000) may generate a tuple (B) based on the generated triple ordered pair and store the generated tuple (B) in the memory unit (2500). At this time, the weight and damping factor assigned to the tuple (B) may be 10 and 0.7, respectively.

[0267] As illustrated in FIG. 22, at time T2, the clone (2000) may receive information “like kpop music” from the user (110) through a predetermined interaction. At this time, the user (110) may be the “ ” in the tuple (B) created at time T1. The clone (2000) may generate a triple ordered pair (John, like, kpop) (C) based on the received information. At this time, the generated triple ordered pair (C) may be composed of the same subject, predicate, and object as the triple ordered pair (A) or the triple ordered pair included in the tuple (B) created at time T1.

[0268] As illustrated in FIG. 22, in response to the generation of identical triple ordered pairs (A, C) at time points T1 and T2, the clone (2000) can generate an updated tuple (D) by updating the damping factor of the tuple (B) stored in the memory unit (2500). Specifically, the damping factor of the updated tuple (D) can be changed from the existing 0.7 to 0.63, which is 10% lower.

[0269] Additionally, the information obtained from user (110) (John) at time T2 can be evaluated as having high reliability by the user (110) providing information about himself. Accordingly, in response to the triple ordered pair (C) identical to the triple ordered pair (A) generated at time T1 being repeatedly generated at time T2 with high reliability, the clone (2000) can generate an updated tuple (D) by updating the weight of the tuple (B). Specifically, the weight of the updated tuple (D) can be changed from the existing 10 to 11, which is an increase of 10%.

[0270] Meanwhile, FIGS. 18 to 22 are merely examples for convenience of explanation, and the present invention is not limited thereto.

[0271]

[0272] According to one embodiment of the present invention, by storing context information in the memory unit (2500) in the form of triple ordered pairs or tuples, the relationship between individual information can be clearly defined based on the structure of subject, predicate, and object. As described above, the normalized data structure can provide the advantage of facilitating mutual referencing and long-term memory management among multiple clones (2000) while minimizing the use of memory resources.

[0273] According to one embodiment of the present invention, a clone (2000) stores context information in a memory unit (2500) in the form of a tuple generated by applying a predetermined weight and attenuation factor to a triple ordered pair, thereby enabling changes in importance over time or memory reinforcement due to repeated input. Accordingly, the clone (2000) can automatically erase unnecessary information while maintaining important information for a long period, thus enabling more accurate and personalized context reflection when generating a response. Furthermore, the above-described triple ordered pair or tuple-based storage structure prevents duplicate storage of the same data, thereby increasing search and inference efficiency, and can maintain consistency in the process of sharing or integrating context information in a multi-agent environment.

[0274]

[0275] Examples

[0276] Social timeline service

[0277] FIG. 23 is a diagram showing an embodiment of a social timeline service by a multi-artificial intelligence agent system according to the present invention.

[0278] According to one embodiment of the present invention, a clone (2000) can interact with at least one other clone (2000) and / or at least one user (110). Specifically, the clone (2000) can obtain context information (2520) related to at least one other clone (2000) and / or at least one user (110) through the above-described interaction.

[0279] As illustrated in FIG. 23, a clone (2000) can generate at least one post (5100) based on acquired context information (2520). Specifically, the post (5100) may be a result expressing the content of an interaction performed by the clone (2000) in relation to at least one other clone (2000) and / or at least one user (110).

[0280] As illustrated in FIG. 23, at least one post (5100) created by a clone (2000) can be posted on a social timeline (5000). Specifically, the social timeline (5000) is for posting the post (5100) created by the clone (2000) and may be a virtual logical space implemented within a local server (200) or separately from the local server (200). A user (110) or a clone (2000) can access the social timeline (5000) and view at least one post (5100) posted on the social timeline (5000).

[0281]

[0282] Music recommendation service

[0283] FIG. 24 is a drawing showing an embodiment of a music recommendation service provided by a clone according to the present invention.

[0284] According to one embodiment of the present invention, a clone (2000) may provide a music recommendation service that recommends a certain music to a user (110) at the request of the user (110) or voluntarily, and provides a certain link to check the recommended music. Specifically, the clone (2000) may select music desired by the user (110) based on context information (2520) stored in a memory unit (2500). At this time, the clone (2000) may access at least some of the external resources described above through a network unit (2100) and may perform operations such as music search and link identification using the external resources.

[0285] Hereinafter, an embodiment of a music recommendation service will be described in detail with reference to FIG. 24. Meanwhile, FIG. 24 is intended only to illustrate an embodiment of a music recommendation service, and the present invention is not limited thereto, and the music recommendation service according to the present invention may be provided in various forms based on various external resources.

[0286]

[0287] As illustrated in FIG. 24, the clone (2000) may receive a specific request for a music recommendation service from the user (110). For example, the clone (2000) may receive a request for “the latest kpop” from the user (110) (see ① in FIG. 24).

[0288] As illustrated in FIG. 24, the clone (2000) can recommend certain music in response to a request from a user (110). Specifically, the clone (2000) can generate a response based on contextual information (2520) related to the user (110), and may utilize external resources such as web search engines and large-scale language models. Here, the contextual information (2520) may include information regarding the user's (110) preferences or music recommendation services provided in the past, but is not limited thereto. For example, the clone (2000) may, in response to a request from the user (110) (see ① in FIG. 24), recommend certain<kpop list> It can provide (see ② in Fig. 24).

[0289] As illustrated in FIG. 24, the clone (2000) may receive a request from the user (110) requesting the search for recommended music. For example, the clone (2000) may receive a request from the user (110) for “song in youtube” (see ③ in FIG. 24).

[0290] As illustrated in FIG. 24, the clone (2000) can search for recommended music in response to a request from the user (110) and provide the search results in the form of a connection link, etc. Specifically, the clone (2000) can utilize external resources such as YouTube and a large-scale language model. For example, the clone (2000) can respond to a request from the user (110) (see ③ in FIG. 24) and a predetermined<Youtube link> It can provide (see ④ in Fig. 24).

[0291]

[0292] Fashion coordinator service

[0293] FIG. 25 is a drawing showing an embodiment of a fashion coordinator service provided by a clone according to the present invention.

[0294] According to one embodiment of the present invention, a clone (2000) may provide information on a predetermined fashion trend to a user (110) either upon the request of the user (110) or voluntarily, and may provide a fashion coordinator service that recommends an outfit suitable for a specific event based on the provided fashion trend. Specifically, the clone (2000) may select an outfit desired by the user (110) based on context information (2520) stored in a memory unit (2500). At this time, the clone (2000) may access at least some of the external resources described above through a network unit (2100) and may perform operations such as searching for fashion trends and recommending outfits using the external resources.

[0295] Hereinafter, an embodiment of a fashion coordinator service will be described in detail with reference to FIG. 25. Meanwhile, FIG. 25 is intended only to illustrate an embodiment of a fashion coordinator service, and the present invention is not limited thereto, and the fashion coordinator service according to the present invention may be provided in various forms based on various external resources.

[0296]

[0297] As illustrated in FIG. 25, the clone (2000) may receive a specific request from the user (110) for searching for fashion trends. For example, the clone (2000) may receive a request from the user (110) saying “Find me the latest fashion style” (see ① in FIG. 25).

[0298] As illustrated in FIG. 25, the clone (2000) can search for and provide information on certain fashion trends in response to a request from a user (110). Specifically, the clone (2000) can utilize external resources such as web search engines, image analysis tools, and large-scale language models.

[0299] More specifically, the clone (2000) can provide information about fashion trends in text form by utilizing a web search engine. For example, the clone (2000) can provide a response “T-shirt and blue jean” in response to a request from a user (110) (see ① in FIG. 25) (see ② in FIG. 25).

[0300] More specifically, the clone (2000) can provide information about fashion trends in the form of images by utilizing a web search engine. For example, the clone (2000) responds to a request from a user (110) (see ① in FIG. 25)<Photo of the latest fashion model> It can provide (see ③ in FIG. 25). At this time, the clone (2000) can provide the analysis results for the provided image by utilizing an image analysis tool. For example, the clone (2000)<Photo of the latest fashion model> Along with the response<Fashion analysis of the photo> A response can be provided (see ③ in Fig. 25).

[0301] As illustrated in FIG. 25, the clone (2000) may receive a request from the user (110) for an outfit recommendation based on a searched fashion trend. For example, the clone (2000) may receive a request from the user (110) for “fashion for a date tonight” (see ④ in FIG. 25).

[0302] As illustrated in FIG. 25, the clone (2000) can recommend an outfit to the user (110) that matches the discovered fashion trends in response to the user's (110) request. Specifically, the clone (200) can generate a response based on contextual information (2520) related to the user (110). Here, the contextual information (2520) may include, but is not limited to, the user's (110) preferences, user profile (including height or weight, etc.), or information about the clothing owned by the user. For example, the clone (2000) can provide a response of “is “”””””” in response to the user's (110) request (see ④ in FIG. 25) (see ⑤ in FIG. 25).

[0303] Additionally, the clone (2000) can utilize an image generation tool to provide a photo of the user (110) wearing the recommended outfit in the form of an image. For example, the clone (2000) can provide a photo of the outfit in the form of an image along with a response that says “is a photo of the suggested style” (see ⑤ in FIG. 25).

[0304] In addition, the clone (2000) may provide additional responses such as “white jacket as tonight is colder than other days” considering the weather conditions. Alternatively, the clone (2000) may use an online shopping platform to provide the user (110) with information on purchasing clothing that matches fashion trends.

[0305]

[0306] Accommodation reservation service

[0307] FIG. 26 is a drawing showing an embodiment of an accommodation reservation service provided by a clone according to the present invention.

[0308] According to one embodiment of the present invention, a clone (2000) may provide a lodging reservation service that recommends a specific lodging to a user (110) upon the user's (110) request or voluntarily, and provides reservation information for the recommended lodging. Specifically, the clone (2000) may select a lodging desired by the user (110) based on context information (2520) stored in a memory unit (2500). At this time, the clone (2000) may access at least some of the external resources described above through a network unit (2100) and may perform operations such as searching for lodging and providing reservation information using the external resources.

[0309] Hereinafter, an embodiment of an accommodation reservation service will be described in detail with reference to FIG. 26. Meanwhile, FIG. 26 is intended only to illustrate an embodiment of an accommodation reservation service, and the present invention is not limited thereto, and the accommodation reservation service according to the present invention may be provided in various forms based on various external resources.

[0310]

[0311] As illustrated in FIG. 26, the clone (2000) may receive a specific request for accommodation recommendations from the user (110). For example, the clone (2000) may receive a request from the user (110) such as “hotels in Silicon Valley such as Hyatt” (see ① in FIG. 26).

[0312] As illustrated in FIG. 26, the clone (2000) can recommend a specific accommodation in response to a request from a user (110). Specifically, the clone (2000) can generate a response based on contextual information (2520) related to the user (110), and in doing so, it may utilize external resources such as web search engines and large-scale language models. Here, the contextual information (2520) may include information regarding the user's (110) preferences or travel destinations, but is not limited thereto. For example, the clone (2000) may, in response to a request from the user (110) (see ① in FIG. 26), recommend a specific It can provide (see ② in Fig. 26).

[0313] As illustrated in FIG. 26, the clone (2000) may receive a request from the user (110) requesting reservation information for a recommended accommodation. For example, the clone (2000) may receive a request from the user (110) saying "a room available, and make a reservation" (see ③ in FIG. 26).

[0314] As illustrated in FIG. 26, the clone (2000) can search for reservation information of recommended accommodations in response to a request from the user (110) and provide the searched results to the user (110). Specifically, the clone (2000) can utilize external resources such as accommodation reservation platforms and large-scale language models. For example, the clone (2000) responds to a request from the user (110) (see ③ in FIG. 26)<List of room types> or provide,<Connects to a booking site> Services can be provided (see ④ in Fig. 26).

[0315]

[0316] User matching service

[0317] FIG. 27 is a drawing illustrating an embodiment of a user matching service provided by a clone according to the present invention.

[0318] According to one embodiment of the present invention, a clone (2000) may provide a user matching service that arranges a meeting between a user (110) and another user (110) either at the request of a user (110) or voluntarily. Specifically, the clone (2000) may perform a matching operation between a user (110) and another user (110) desired by the user (110) based on context information (2520) stored in a memory unit (2500). At this time, the clone (2000) may access at least some of the external resources described above through a network unit (2100) and may perform an operation to search for a user (110) and / or a clone (2000) using the external resources.

[0319] Hereinafter, an embodiment of a user matching service will be described in detail with reference to FIG. 27. Meanwhile, FIG. 27 is intended only to illustrate an embodiment of a user matching service, and the present invention is not limited thereto, and the user matching service according to the present invention may be provided in various forms based on various external resources.

[0320]

[0321] As illustrated in FIG. 27, the clone (2000) may receive a specific request from the user (110) for a meeting place recommendation service. For example, the clone (2000) may receive a request from the user (110) such as "List dating places" or "List romantic outdoor places" (see ① in FIG. 27).

[0322] As illustrated in FIG. 27, the clone (2000) can recommend a predetermined meeting place in response to a request from a user (110). Specifically, the clone (2000) can generate a response based on contextual information (2520) related to the user (110), and may utilize external resources such as a large-scale language model. Here, the contextual information (2520) may include information regarding the user's (110) preferences, personality, or individuality, but is not limited thereto. For example, the clone (2000) may, in response to a request from the user (110) (see ① in FIG. 27), recommend a predetermined<List of dating places> or<List of places such as beaches with beautiful sunset> It can provide (see ② in Fig. 27).

[0323] As illustrated in FIG. 27, the clone (2000) may receive a request from the user (110) requesting the search for another clone (2000) created or trained by another user (110) desired by the user (110). For example, the clone (2000) may receive a request from the user (110) stating "clones trained with female born in 1995" (see 3 in FIG. 27).

[0324] As illustrated in FIG. 27, the clone (2000) may search for other clones (2000) created or trained by other users (110) desired by the user (110) in response to a request from the user (110), and provide the search results to the user (110). Specifically, the clone (2000) may generate a response based on contextual information (2520) related to the user (110), and may utilize external resources such as dating platforms and large-scale language models. For example, the clone (2000) may respond to a request from the user (110) (see ③ in FIG. 27) and a predetermined<List of clones trained with female born in 1995> It can provide (see ④ in Fig. 27).

[0325] Meanwhile, the information about the user (110) provided to the clone (2000) from a dating platform, etc., may be subject to homomorphic encryption. Additionally, during the interaction between clones (2000), the information about the user (110) associated with the actual clone (2000) may be utilized without being decrypted. Therefore, the leakage of the user's (110) information during the interaction can be prevented.

[0326] As illustrated in FIG. 27, the user (110) may select at least one of the clones (2000) provided by the clone (2000) and may have the clone (2000) interact with the selected clone (2000). For example, the first clone (2000-1) may receive requests from the first user (110-1) such as "Choose Clone 2" and "Start dating with Clone 2," and additionally receive a request such as "Continue dating with Clone 2" (see ⑤ in FIG. 27). The first clone (2000-1) may continuously interact with the second clone (2000-2) in response to the requests of the first user (110-1) (see ⑥ in FIG. 27). At this time, the second clone (2000-2) may be an agent created and / or trained by the second user (110-2).

[0327] As illustrated in FIG. 27, the user (110) may have the clone (2000) send a meeting request to another clone (2000). For example, the first clone (2000-1) may receive a request from the first user (110-1) that says "if Clone 2 agrees on a meeting" (see ⑦ in FIG. 27). The first clone (2000-1) may send a meeting request to the second clone (2000-2) in response to the request from the first user (110-1). Upon receiving the meeting request, the second clone (2000-2) may provide the second user (110-2) with information about the first clone (2000-1) (and the first user (110-1)) along with the fact that the meeting request was received. The second user (110-2) can decide whether to accept the meeting request of the first user (110-1) based on the information provided by the second clone (2000-2).

[0328]

[0329] Event Recording Service

[0330] FIG. 28 is a drawing illustrating an embodiment of an event recording service provided by a clone according to the present invention.

[0331] According to one embodiment of the present invention, a clone (2000) may provide an event recording service that records the daily life or experiences of a user (110) either at the request of the user (110) or voluntarily. Specifically, the clone (2000) may generate a recording result based on context information (2520) stored in a memory unit (2500) and information about the daily life or experiences provided by the user (110). At this time, the clone (2000) may access at least some of the external resources described above through a network unit (2100), and may perform an operation to analyze the information provided by the user (110) using the external resources and generate a recording result.

[0332] Hereinafter, an embodiment of an event recording service will be described in detail with reference to FIG. 28. Meanwhile, FIG. 28 is intended only to illustrate an embodiment of an event recording service, and the present invention is not limited thereto, and the event recording service according to the present invention may be provided in various forms based on various external resources.

[0333]

[0334] As illustrated in FIG. 28, the clone (2000) may receive information related to the user's (110) daily life or experiences from the user (110). For example, the information described above may be provided in the form of images or videos, but is not limited thereto (see ① in FIG. 28).

[0335] As illustrated in FIG. 28, the clone (2000) can generate a predetermined record result based on information received from the user (110) and provide it to the user (110). Specifically, the clone (2000) can analyze the information received from the user (110) by utilizing an image analysis tool, a video analysis tool, and a large-scale language model, and generate a response based on the analysis result and contextual information (2520) related to the user (110). Here, the contextual information (2520) may include information such as the time when an interaction with the user (110) took place or a place related to the interaction, but is not limited thereto. For example, the clone (2000) to the user (110) It can provide.

[0336]

[0337] Coordination of conversations between multiple users and clones

[0338] Agent clone

[0339] FIG. 29 is a diagram illustrating an embodiment of a proxy discussion service provided by a proxy clone according to the present invention.

[0340] According to one embodiment of the present invention, a proxy clone (6000) can conduct a conversation or discussion among multiple people on behalf of at least one user (110) and at least one clone (2000). The user (110) or the clone (2000) may transmit content that they wish to speak about in relation to the ongoing conversation or discussion to the proxy clone (6000), and the proxy clone (6000), upon receiving the content, may organize or adapt the content and provide the content on the channel (1000) where the conversation or discussion is conducted on behalf of the user (110) or the clone (2000).

[0341] For example, as illustrated in FIG. 29, a discussion between multiple users (110) and multiple clones (2000) may take place on the channel (1000). Specifically, the first proxy clone (6000-1) may represent some of the discussion participants, and the second proxy clone (6000-2) may represent the remaining participants. In practice, the subjects speaking on the channel (1000) may be the first proxy clone (6000-1) and the second proxy clone (6000-2).

[0342]

[0343] Operation to determine utterance time

[0344] FIG. 30 is a diagram showing the operation of determining the firing time of a clone according to the present invention.

[0345] According to one embodiment of the present invention, when interaction occurs between a plurality of clones (2000) on a channel (1000), the clone (2000) can determine its utterance time through a predetermined method. Specifically, the clone (2000) can predict an expected waiting time based on the content of the interaction occurring on the channel (1000). Additionally, the clone (2000) can continue uttering after waiting for the predicted expected waiting time.

[0346] As illustrated in FIG. 30, the clone (2000) can calculate an expected waiting time by using a predetermined prompt as input to a large-scale language model. For example, the prompt may be "Read carefully the given conversation and predict the likelihood that you will be the next speaker in the conversation. The likelihood value should be between 0 and 1.0," but is not limited thereto. Specifically, the clone (2000) can calculate the probability that it will be the next speaker and calculate an expected waiting time based on the calculated probability. After waiting for the calculated expected waiting time, if the clone (2000) determines that the time for speaking has arrived, it may post a predetermined speech indicator on the channel (1000). For example, the speech indicator may be a period (""), but is not limited thereto. After the speech indicator is posted, the clone (2000) may proceed with speaking based on the large-scale language model.

[0347]

[0348] Global Prompt

[0349] According to one embodiment of the present invention, a plurality of clones (2000) interacting on a channel (1000) may be subject to a predetermined global prompt. Specifically, the global prompt is a prompt that is commonly applied to a plurality of clones (2000) implemented within the channel (1000), and may be intended to impose a predetermined rule on the behavior of the clones (2000). A clone (2000) to which a global prompt is applied is configured to act under the control of a predetermined rule defined by the global prompt, and interaction between the clones (2000) may also take place under the aforementioned rule.

[0350] For example, a user (110) can open a channel (1000) named "Clones for planning to travel to Tokyo". At this time, the user (110) can apply a global prompt on the channel (1000) stating "AI agents must speak in Japanese from now on". Accordingly, all clones (2000) on the channel (1000) can be configured to interact exclusively in Japanese. However, they are not limited to this.

[0351] Meanwhile, the global prompt is applied to the channel (1000), and the clone (2000) implemented within the channel (1000) may be configured to be under the control of the global prompt. However, the global prompt does not change the attributes of the clone (2000) itself, and if the clone (2000) is implemented in a channel (1000) other than the aforementioned channel (1000), it may no longer be affected by the aforementioned global prompt.

[0352]

[0353] scenario

[0354] FIG. 31 is a diagram showing that a scenario is extracted from an interaction between at least one user and / or at least one clone according to the present invention.

[0355] As illustrated in FIG. 31, a user (110) can interact with at least one clone (2000) (hereinafter referred to as a target interaction (T)). Specifically, the target interaction (T) may consist of at least one sub-action. For example, the target interaction (T) may consist of a first interaction (t1) through a Nth interaction (tN).

[0356] At this time, the user (110) can extract a scenario (S) corresponding to a target interaction (T) through a predetermined method. Specifically, the scenario (S) may include all information regarding the content and progress method of the target interaction (T). Here, the scenario (S) may include at least one step corresponding to each of at least one sub-action constituting the target interaction (T). For example, the scenario (S) may be composed of a first step (s1) to a Nth step (sN), and the first step (s1) to the Nth step (sN) may each correspond to a first interaction (t1) to a Nth interaction (tN).

[0357] FIG. 32 is a diagram showing that a scenario is applied to the interaction between at least one user and / or at least one clone according to the present invention.

[0358] As illustrated in FIG. 32, an interaction (hereinafter referred to as the target interaction (T') distinct from the aforementioned target interaction (T) may occur between a user (110) and at least one clone (2000). At this time, at least one clone (2000) may be subjected to a predetermined scenario (S) extracted from the target interaction (T). The scenario (S) may be applied to the clone (2000) in the form of a predetermined prompt, but is not limited thereto. At least one clone (2000) subjected to the scenario (S) may be configured to interact with the user (110) such that the target interaction (T') proceeds based on content and a method of progress identical or similar to that of the target interaction (T). Specifically, the fact that the target interaction (T') and the target interaction (T) are identical or similar may mean that each of the first interaction (t1') to the Nth interaction (tN'), which is a sub-action of the target interaction (T'), is identical or similar to the first interaction (t1) to the Nth interaction (tN), which is a sub-action of the target interaction (T).

[0359] According to one embodiment of the present invention, a scenario (S) can be applied to at least one clone (2000) in stages. For example, the first stage (s1) to the Nth stage (sN) constituting the scenario (S) can each be applied sequentially to at least one clone (2000) during the first interaction (t1') to the Nth interaction (tN') process. Specifically, when each sub-stage included in the scenario (S) is applied sequentially to the clone (2000), the similarity between the target interaction (T') and the target interaction (T) can be increased compared to when the entire scenario (S) is applied to the clone (2000) in the form of a single prompt.

[0360]

[0361] Nomadic AI Agent

[0362] FIG. 33 is a diagram showing a nomadic artificial intelligence agent implemented in a nomadic artificial intelligence environment according to the present invention.

[0363] According to one embodiment of the present invention, a clone (2000) generated at a local server (200) may be implemented in a space other than the local server (200) in a predetermined manner. Specifically, the clone (2000) may be implemented in a predetermined local device distinct from the local server (200), and in this case, the predetermined context information (2520) held by the clone (2000) in relation to the user (110) may be maintained. More specifically, the local device may refer to a device or system equipped with computational resources and memory resources capable of calling or executing the clone (2000) so that the clone (2000) is implemented temporarily or permanently. For example, the local device may include, but is not limited to, a vehicle, an automotive electronic system, a smart device, a tablet PC, a wearable device, an AI speaker, an IoT control device, a home appliance, a smart city, a smart home, a smart office, a drone, and a robot. The clone (2000) can be executed in a local device environment as described above in a predetermined manner, away from the cloud environment of the local server (200), and the clone (2000) can maintain the context information obtained from the local server (200) as is. Additionally, the clone (2000) can be configured to provide a response to a request from a user (110) by utilizing a predetermined tool available in the local device environment. That is, the clone (2000) can switch the execution environment from the local server (200) to the local device while maintaining identity in terms of context information (2520).

[0364] In the specification of the present invention, the local device described above, in which a clone (2000) other than the local server (200) is implemented, may be interchangeably referred to as a Nomadic AI environment (7000). Additionally, the clone (2000) implemented in the Nomadic AI environment (7000) described above may be interchangeably referred to as a Nomadic AI agent (7500).

[0365] Hereinafter, the nomadic artificial intelligence environment (7000) and the nomadic artificial intelligence agent (7500) according to the present invention will be described in detail with reference to FIG. 33.

[0366]

[0367] As illustrated in FIG. 33, the nomadic artificial intelligence environment (7000) may include a nomadic artificial intelligence execution module (7100). Specifically, the nomadic artificial intelligence execution module (7100) may be configured to implement or execute a clone (2000) corresponding to the received context information (2520) in the nomadic artificial intelligence environment (7000) based on the context information (2520) received from the local server (200).

[0368] According to one embodiment of the present invention, the nomadic artificial intelligence execution module (7000) may include a network unit (7110), a processor unit (7130), and a memory unit (7150).

[0369] Specifically, the nomadic artificial intelligence execution module (7000) can receive context information (2520) constituting a clone (2000) implemented on the local server (200) from the local server (200) through the network unit (7110). Alternatively, the nomadic artificial intelligence execution module (7000) can connect to the local server (200) through the network unit (7110) and obtain context information (2520) constituting a clone (2000) implemented on the local server (200).

[0370] Specifically, the processor unit (7130) can implement or execute a nomadic artificial intelligence agent (7500) in a nomadic artificial intelligence environment (7000) based on context information (2520) received or acquired from a local server (200).

[0371] As illustrated in FIG. 33, a nomadic artificial intelligence agent (7500) implemented by a nomadic artificial intelligence execution module (7100) may include a network unit (7510), a processor unit (7530), and a memory unit (7550). In the same manner as described above with respect to a clone (2000) implemented in a local server (200), the nomadic artificial intelligence agent (7500) may also be a logical entity executed in a virtual logical space implemented in a nomadic artificial intelligence environment (7000). At this time, the network unit (7510), processor unit (7530), and memory unit (7550) of the nomadic artificial intelligence agent (7500) may be configured to operate by allocating at least a portion of the resources of the network unit (7110), processor unit (7130), and memory unit (7150) of the nomadic artificial intelligence execution module (7100), respectively.

[0372] According to one embodiment of the present invention, the memory unit (7550) of the nomadic artificial intelligence agent (7500) may include context information (2520) obtained from the local server (200) by the nomadic artificial intelligence execution module (7100). Specifically, by the nomadic artificial intelligence agent (7500) possessing the same context information (2520) as the clone (2000) created and implemented on the local server (200), identity may be maintained in the relationship between the nomadic artificial intelligence agent (7500) and the clone (2000).

[0373] According to one embodiment of the present invention, the processor unit (7530) of the nomadic artificial intelligence agent (7500) may be configured to generate and provide a predetermined response in response to a request from a user (110). Specifically, the processor unit (7530) may utilize a predetermined large-scale language model provided in the memory unit (7150) of the nomadic artificial intelligence execution module (7100) or in the nomadic artificial intelligence environment (7000). More specifically, the aforementioned large-scale language model may be a model specialized for the nomadic artificial intelligence environment (7000), and may be lightweight to a predetermined degree compared to the large-scale language model utilized by the clone (2000) implemented in the local server (200). For example, if the local device corresponding to the nomadic artificial intelligence environment (7000) is a vehicle, the aforementioned large-scale language model may be a model learned based on training data, etc., for controlling the operation of the vehicle, but is not limited thereto.

[0374] According to one embodiment of the present invention, the network section (7110) of the nomadic artificial intelligence execution module (7100) may include an MCP interface (Model Context Protocol interface) (7111). Specifically, the MCP interface (7111) may be configured to perform the function of a standard connection protocol for accessing a predetermined operation module (7300) for controlling the operation of the nomadic artificial intelligence environment (7000). The nomadic artificial intelligence execution module (7100) may access the aforementioned operation module (7300) through the MCP interface (7111) and control the operation of the nomadic artificial intelligence environment (7000) or obtain information regarding the operation status of the nomadic artificial intelligence environment (7000). The nomadic artificial intelligence agent (7500) may be configured to access the operation module (7300) of the nomadic artificial intelligence environment (7000) through the MCP interface (7111) by allocating at least a portion of the computational resources of the nomadic artificial intelligence execution module (7100) to operate. For example, if the local device corresponding to the nomadic artificial intelligence environment (7000) is a vehicle, the operation module (7300) may be configured to control the operation of the vehicle's engine, navigation, and heating / cooling system. In this case, the nomadic artificial intelligence agent (7500) may control the operation of the vehicle's engine, navigation, and heating / cooling system through the MCP interface (7111), or obtain information regarding the operating status of the engine in the vehicle, the speed of the vehicle, the location of the vehicle, and the heating / cooling status in the vehicle. However, it is not limited thereto. Thus, the nomadic artificial intelligence agent (7500) can generate a response to the user's (110) request based on context information (2520) related to the user (110) and information about the operating state of the nomadic artificial intelligence environment (7000).

[0375]

[0376] Although the present invention described above has been explained with reference to the embodiments illustrated in the drawings, this is merely illustrative, and those skilled in the art will understand that various modifications and variations of the embodiments are possible therefrom. However, such modifications should be considered to be within the technical scope of protection of the present invention. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the appended claims.

[0377] [Explanation of the symbol]

[0378] 10: Multi-AI Agent System

[0379] 100 : User terminal

[0380] 110 : User

[0381] 110-1 : First user

[0382] 110-2 : Second user

[0383] 200 : Local Server

[0384] 210 : Network Department

[0385] 230: Processor section

[0386] 250 : Memory section

[0387] 251 : Prompt storage

[0388] 252 : Sub-dialogue information

[0389] 253 : Image storage section

[0390] 1000 : Channel

[0391] 2000: Clone

[0392] 2000-1 : 1st Clone

[0393] 2000-2 : Second Clone

[0394] 2100 : Network Department

[0395] 2300: Processor section

[0396] 2500 : Memory section

[0397] 2510 : Prompt information

[0398] 2520 : Contextual Information

[0399] 3000 : Orchestrator Clone

[0400] 4000 : Conversation Manager Clone

[0401] 5000 : Social Timeline

[0402] 5100 : Post

[0403] 6000 : Agent Clone

[0404] 6000-1 : 1st Delegate Clone

[0405] 6000-2 : Second Agent Clone

[0406] T: Target Interaction

[0407] t1 to tN: First interaction to Nth interaction

[0408] S: Scenario

[0409] s1 to sN: Steps 1 to N

[0410] T' : Object interaction

[0411] t1' to tN': First interaction to Nth interaction

[0412] 7000: Nomadic AI Environment

[0413] 7100: Nomadic AI Execution Module

[0414] 7110 : Network Section

[0415] 7111: MCP Interface

[0416] 7130: Processor section

[0417] 7150: Memory section

[0418] 7300 : Operation Module

[0419] 7500 : Nomadic AI Agent

[0420] 7510 : Network Section

[0421] 7530: Processor section

[0422] 7550: Memory section

Claims

In a computing device for building a Multi Artificial Intelligence (AI) Agent environment, A network unit for transmitting / receiving information in relation to at least one user terminal and at least one external resource; A processor unit for creating at least one clone, creating at least one channel which is a logical space in which the at least one clone created is logically implemented, and calling the at least one clone in the at least one channel created; and A memory unit for storing information about the at least one clone and the at least one channel; including, A computing device for building a multi-artificial intelligence agent environment. In Article 1, The above external resource is at least one resource selected from the group including storage service resources, social network service resources, search engine service resources, and artificial intelligence model resources, A computing device for building a multi-artificial intelligence agent environment. In Article 1, The above processor unit is, A step of enabling the network unit to receive at least one piece of information selected from a group including a clone generation prompt and biological information from the user terminal; and A step of generating the clone based on at least one of the received information; Configured to perform, A computing device for building a multi-artificial intelligence agent environment. In Paragraph 3, The above memory unit includes a prompt storage unit for storing at least one prompt for generating the clone, and The above processor unit is, A step of outputting a final prompt using the received clone generation prompt and at least one prompt stored in the prompt storage unit as inputs for a Large Language Model (LLM); and A step of generating the clone based on the final prompt output above; Configured to perform further, A computing device for building a multi-artificial intelligence agent environment. In Article 4, The final prompt output above is stored in the prompt storage unit, A computing device for building a multi-artificial intelligence agent environment. In Paragraph 3, The above processor unit is, A step of extracting biological features using the received biological information as input to a feature extraction model; A step of extracting a persona using the extracted biological characteristics as input to a persona extraction model; A step of outputting the clone generation prompt using the extracted persona as input to a Large Language Model (LLM); and A step of generating the clone based on the clone generation prompt output above; Configured to perform further, A computing device for building a multi-artificial intelligence agent environment. In Paragraph 3, The above memory unit includes an image storage unit for storing information about a sequence pair of at least one profile image and a corresponding image space vector. The above processor unit is, A step of causing the network unit to receive a profile image generation prompt from the user terminal; A step of outputting a prompt space vector corresponding to the profile image generation prompt using the received profile image generation prompt as input to a vector embedding model; and A step of selecting at least one profile image for the generated clone from the image storage unit through matching between the output prompt space vector and the ordered pair; Configured to perform further, A computing device for building a multi-artificial intelligence agent environment. In Article 1, The above clone is, A network unit for transmitting / receiving information in relation to at least one user terminal, the computing device, and at least one other clone; A processor unit for generating a response to a user request received from the above-mentioned user terminal; and A memory unit for storing information collected from the above-mentioned at least one user terminal and the above-mentioned at least one channel—hereinafter referred to as context information—; including, A computing device for building a multi-artificial intelligence agent environment. In Article 8, The network unit, the processor unit, and the memory unit of the above clone are each configured to operate by being allocated at least a portion of the resources of the network unit, the processor unit, and the memory unit of the computing device, respectively. A computing device for building a multi-artificial intelligence agent environment. In Article 8, The processor part of the above computing device is, Configured to call at least one other clone to the channel in response to a call request of the clone implemented in the channel, A computing device for building a multi-artificial intelligence agent environment. In Article 8, The memory portion of the above clone includes prompt information for generating an input prompt corresponding to the user request, and The processor part of the above clone is, A step of outputting a coordination prompt using the above user request, the above prompt information, and the above context information as inputs to a first Large Language Model (LLM); and A step of generating the response to the user request using the second large-scale language model of the adjustment prompt output above as input; Configured to perform, A computing device for building a multi-artificial intelligence agent environment. In Article 8, The processor part of the above clone is, A step of causing the network portion of the clone to receive at least one link from the external resource in response to the above user request; A step of verifying the validity of at least one received link; and A step of generating the response to the user request based on the validated link among the at least one link; Configured to perform, A computing device for building a multi-artificial intelligence agent environment. In a computing device for building a Multi Artificial Intelligence (AI) Agent environment, A network unit for transmitting / receiving information in relation to at least one user terminal and at least one external resource; A processor unit for creating at least one clone, creating at least one channel which is a logical space in which the at least one clone created is logically implemented, and calling the at least one clone in the at least one channel created; and A memory unit for storing information about the at least one clone and the at least one channel; including, A computing device for building a multi-artificial intelligence agent environment. In Article 13, The above clone is, A network unit for transmitting / receiving information in relation to at least one user terminal, the computing device, and at least one other clone; A processor unit for generating a response to a user request received from the above-mentioned user terminal; and A memory unit for storing information collected from the above-mentioned at least one user terminal and the above-mentioned at least one channel—hereinafter referred to as context information—; including, A computing device for building a multi-artificial intelligence agent environment. In Article 14, The above processor unit is configured to generate a triple ordered pair (Triple) corresponding to the context information based on the collected context information, and The above context information is stored in the memory unit in the form of the generated triple ordered pair, A computing device for building a multi-artificial intelligence agent environment. In Article 15, The processor unit is configured to generate the response to the user request based on the context information stored in the form of the triple ordered pair. A computing device for building a multi-artificial intelligence agent environment. In Article 15, The processor unit is configured to determine a weight and a decay factor for the triple ordered pair, and to generate a tuple by combining the triple ordered pair and the weight and decay factor determined in relation to the triple ordered pair. A computing device for building a multi-artificial intelligence agent environment. In Article 17, The above context information is stored in the memory unit in the form of the generated tuple, A computing device for building a multi-artificial intelligence agent environment. In Article 17, The weights of the above tuple are expressed by [Equation 1], A computing device for building a multi-artificial intelligence agent environment. [Mathematical Formula 1] (Here, W T : The weights for the triple ordered pairs at time T, W T=0 : Initial value of the weight of the triple ordered pair at time T0 (the time when the triple ordered pair was first generated), T: Time elapsed since the time when the above triple ordered pair was first created, Df: The damping factor for the above triple ordered pair) In Article 17, The above processor unit is, A step of generating a first triple ordered pair based on first context information, determining weights and damping factors for the generated first triple ordered pair, and generating a first tuple by combining the first triple ordered pair and the weights and damping factors determined in relation to the first triple ordered pair; A step of generating a second triple ordered pair based on second context information; and A step of updating the damping factor of the first tuple in response to the determination that the first triple ordered pair is identical to the generated second triple ordered pair; Configured to perform, A computing device for building a multi-artificial intelligence agent environment. In Article 20, The above processor unit is, A step of evaluating the reliability of the above second context information; and A step of updating the weight of the first tuple in response to the determination that the reliability is greater than or equal to a predetermined threshold; Configured to perform further, A computing device for building a multi-artificial intelligence agent environment. In an artificial intelligence agent implemented on a server for building a Multi Artificial Intelligence (AI) Agent environment, A network unit for transmitting / receiving information in relation to at least one user terminal, the server, and at least one other artificial intelligence agent; A processor unit for generating a response to a user's request received from the user terminal by utilizing a predetermined external resource accessible through the server; and A memory unit for storing information obtained based on interaction with at least one user terminal or at least one other artificial intelligence agent—hereinafter referred to as context information; Includes, The above network unit, the above processor unit, and the above memory unit are configured to operate by being allocated computing resources of the server, and The response generated by the processor unit is generated based on the context information. Artificial intelligence agent. In Article 22, The processor unit is configured to generate at least one post related to the context information based on the acquired context information, and to post the generated at least one post on a social timeline built on the server. Artificial intelligence agent. In Article 22, The above processor unit is configured to search for information regarding a certain music as a response to the above user request, provide the searched information to the above user, and provide a certain connection link that can verify the information. Artificial intelligence agent. In Article 22, The processor unit is configured to search for information regarding a predetermined fashion trend as a response to the user request, provide the searched information to the user, and recommend a predetermined outfit to the user based on the information. Artificial intelligence agent. In Article 22, The above processor unit is configured to search for information regarding a predetermined accommodation as a response to the above user request, and to provide the searched information to the user. The above information includes reservation information for the above accommodation, Artificial intelligence agent. In Article 22, The processor unit is configured to search for information about at least one other artificial intelligence agent as a response to the user request, provide the searched information to the user, and perform interaction with the artificial intelligence agent selected by the user among the at least one other artificial intelligence agent. The above information is one to which homomorphic encryption has been applied, Artificial intelligence agent. In Article 22, The processor unit is configured to analyze an image or video provided by the user as a response to the user request, generate a predetermined record result related to the image or video based on the analysis result, and provide the generated record result to the user. Artificial intelligence agent. In Article 22, The processor unit is configured to calculate a predetermined expected waiting time based on the context information and to provide the response to the user request to the user after waiting for the calculated expected waiting time. Artificial intelligence agent. Regarding a method for building a Multi Artificial Intelligence (AI) Agent environment, A step of extracting a predetermined scenario from a target interaction performed between at least one user and at least one artificial intelligence agent; and A step of applying the above extracted scenario to a target interaction performed between the at least one user and the at least one artificial intelligence agent; including, Method for building a multi-AI agent environment. In Article 30, The above target interaction is configured as a sequential combination of at least one sub-interaction, Method for building a multi-AI agent environment. In Article 31, The above scenario is configured as a sequential combination of at least one sub-step, and Each of the above at least one sub-step corresponds to each of the above at least one sub-interaction, Method for building a multi-AI agent environment. In Article 32, The above target interaction is configured as a sequential combination of at least one sub-interaction, and Each of the at least one sub-step constituting the above scenario is sequentially applied to each of the at least one sub-interaction constituting the target interaction, Method for building a multi-AI agent environment. In a nomadic artificial intelligence (AI) agent system, A local server for creating an artificial intelligence agent - hereinafter referred to as a clone -; and A local device for generating a nomadic artificial intelligence agent corresponding to the above clone - hereinafter referred to as the nomadic artificial intelligence environment - ; Includes, The above nomadic artificial intelligence environment is, A nomadic artificial intelligence execution module configured to receive information related to the clone—hereinafter referred to as context information—from the local server, and to create and execute the nomadic artificial intelligence agent based on the received context information; Includes, The above-mentioned nomadic artificial intelligence agent is configured to be operated by being allocated computational resources of the above-mentioned nomadic artificial intelligence execution module and to generate a response to a user request based on the above-mentioned context information. Nomadic AI Agent System. In Article 34, The above-mentioned nomadic artificial intelligence execution module is, Network section including an MCP interface; Includes, The above-mentioned nomadic artificial intelligence agent is configured to access a predetermined operation module included in the nomadic artificial intelligence environment through the above-mentioned MCP interface, Nomadic AI Agent System. In Article 34, The above nomadic artificial intelligence agent is configured to generate the response by utilizing a predetermined large-scale language model stored in the above nomadic artificial intelligence execution module, and The above-mentioned large-scale language model is trained based on training data for controlling the operation of the above-mentioned nomadic artificial intelligence environment, Nomadic AI Agent System.

Images