A generative agent system with long short-term memory mechanism and implementation method thereof

Abstract

The application discloses a kind of generative intelligent agent systems with long short-term memory mechanism and implementation method thereof, the system includes intelligent agent generator, intelligent agent editor, intelligent agent executor, data bus, knowledge base, tool library and algorithm library;Intelligent agent generator is responsible for generating workflow, intelligent agent editor is responsible for publishing workflow, intelligent agent executor is responsible for executing workflow, data bus is responsible for the data management inside workflow, ensure data transmission, storage and application in the process of workflow execution, knowledge base, tool library and algorithm library are responsible for data, tool and algorithm storage respectively.The method is implemented based on the system.The application has the following advantages: considering the high stability of workflow and the flexibility of intelligent agent execution, meeting the diversified user demand while ensuring the execution efficiency and reliability in serious engineering business scenarios;Workflow can be quickly generated or adjusted according to business requirements, and one-key publishing can be used as available services, significantly reducing the design and maintenance threshold.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent generation and application technology, specifically to a generative intelligent agent system with a long short-term memory mechanism and its implementation method. Background Technology

[0002] Currently, general-purpose AI large models are prone to erroneous conclusions when dealing with specialized business problems, especially in the water conservancy and water affairs industry, due to data gaps and the inherent limitations of large model technology (the "large model illusion"). While agent technology, by incorporating components available in large models such as knowledge memory, workflow planning, and tool invocation, can improve the accuracy of question answering by introducing necessary data and autonomously planning behavior, certain uncertainties remain in current agent operations, especially when locally deployed large models (models with hundreds of billions or trillions of parameters) cannot be applied. To address these uncertainties, some existing solutions propose using orchestrated workflow techniques, where the agent's workflow path and invoked tools are manually planned. This reduces adaptability while improving the stability of the large model's runtime, as exemplified by Dify and Coze. These solutions primarily connect pre-defined workflows consisting of multiple steps, with the program following the workflow requirements to invoke the large model and its related tools, thereby completing a complex task.

[0003] However, in complex scenarios, the above solutions have shortcomings: First, workflow orchestration technology requires editing software to arrange workflows, and sometimes requires writing code to complete complex tasks. This requires compound talents with expertise in business, technology, and algorithms to complete workflow orchestration and maintenance, making it difficult to implement in actual engineering projects. Second, workflow orchestration technology is difficult to adapt to all possible business scenarios and has a weak ability to adapt to changes in user needs, which limits its scope of use. Third, existing orchestration is often strongly bound to specific business systems and lacks the ability to decouple resources and accumulate data. Summary of the Invention

[0004] The purpose of this invention is to provide a generative intelligent agent system with a long short-term memory mechanism and its implementation method, which solves the problems existing in the background technology such as the illusion of large models, insufficient support of large models for business scenarios, complex workflow orchestration, difficulty in maintaining intelligent agents, and inability to expand the scope of business.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect, the present invention provides a generative intelligent agent system with a long short-term memory mechanism, comprising:

[0007] The intelligent agent generator is responsible for generating workflows tailored to specific business scenarios based on user needs, textual SOPs stored in the knowledge base, and existing workflows.

[0008] The agent editor is responsible for visually displaying the workflow generated by the agent generator, where users can customize or arrange the various work steps of the agent and preview the execution effect;

[0009] The agent executor is responsible for converting the workflow text generated by the agent generator into an executable program, and for performing execution verification and feedback of execution results.

[0010] The data bus is responsible for data management within the workflow, ensuring data transmission, storage, and application during workflow execution.

[0011] The knowledge base is responsible for storing data input from external sources or data accumulated by the system during runtime. The external input data includes relevant business knowledge, terminology database, SOPs, and examples of generated workflows or SQL query statements.

[0012] The tool library is responsible for registering MCP services, discovering a list of general tools available to agents, and allowing agents to select appropriate tools to complete business functions when generating workflows.

[0013] The algorithm library, with the same execution process as the tool library, integrates independent professional business algorithms designed and developed by the business system.

[0014] Furthermore, the workflow generated by the agent generator can be further edited and published by the agent editor.

[0015] Furthermore, the agent editor can directly edit an agent as defined by a composite actuator protocol, or it can load a workflow generated by the agent generator and submit it to the user for editing.

[0016] Furthermore, the agent editor further edits and publishes the workflow generated by the agent generator, including the following steps:

[0017] A1. Workflow Loading: Load an existing workflow using the agent editor, and expand and edit each specific atomic step in the generated workflow;

[0018] A2. Workflow Creation: Users can manually edit the tasks to be performed and their data dependencies for each work node, and plan the execution order of each atomic node; during editing, users can directly generate a workflow template through the agent generator and make adjustments based on it;

[0019] A3. Workflow Editing: Users can adjust the specific behavior of atomic operations and their output and input data dependencies by setting prompts;

[0020] A4. Workflow Validation and Preview: The user clicks the play button at the starting node in the workflow to execute the workflow until the results of all execution steps are displayed; if there are data dependency errors or the workflow cannot run during execution, the system will report an error and explain the reason for the error.

[0021] Furthermore, in the intelligent agent editor, adjacent atomic nodes belong to different execution steps in the same workflow, working nodes in the same row can be executed in parallel, nodes in different rows are executed serially, data is transmitted between rows through a data bus, and each node obtains data from the data bus according to its own definition.

[0022] Furthermore, in the intelligent agent editor, the execution order of the workflow can be directly adjusted by dragging and dropping atomic node blocks to different rows and columns, and an atomic node can be deleted to adjust the execution steps of the entire workflow.

[0023] Furthermore, the intelligent agent editor includes multiple workflow editor canvases, each containing multiple unconnected workflows. These unconnected workflows can be executed individually or synchronously via the play button on the canvas.

[0024] Furthermore, the agent editor includes multiple workflow editor canvases, each workflow editor canvas containing multiple workflows, or a single workflow, or published as a separate service to form a single agent service.

[0025] Secondly, the present invention provides a method for implementing a generative intelligent agent system with a long short-term memory mechanism, comprising the following steps:

[0026] B1. Users input user questions into a generative intelligent agent system with a long short-term memory mechanism based on their needs;

[0027] B2. The generative intelligent agent system with long short-term memory mechanism determines whether there is a workflow in its system that matches the user's question based on the user's question;

[0028] B3. If there is a workflow in the system that matches the user problem, then the agent executor in the generative agent system with long short-term memory mechanism executes the workflow according to the workflow that matches the user problem.

[0029] B4. If there is no workflow in the system that matches the user problem, the agent generator in the generative agent system with long short-term memory first generates a workflow for the user problem. Then, the agent editor in the generative agent system with long short-term memory visualizes the workflow generated by the agent generator. After that, the agent executor in the generative agent system with long short-term memory executes the workflow generated by the agent generator. Finally, the successfully executed and published workflow generated by the agent generator for the user problem is saved in the agent memory.

[0030] Compared with the prior art, the advantages of the present invention are as follows:

[0031] (1) It balances the high stability of the workflow with the flexibility of the agent execution, and can ensure the execution efficiency and reliability in serious engineering business scenarios while meeting diverse user needs.

[0032] (2) Through the intelligent agent editor, a visual and WYSIWYG workflow orchestration and maintenance method is provided, which enables users to easily update the intelligent agent logic without professional technical background, significantly reducing the design and maintenance threshold; combined with knowledge memory, tool call and autonomous behavior planning, workflows can be quickly generated or adjusted according to business needs, and customized workflows can be published as available services with one click, which can quickly connect to business systems and achieve efficient integration between development and application.

[0033] (3) Relying on the resource system composed of knowledge base, tool library and algorithm library, the AI ​​capabilities and dependent resources are decoupled, so that professional business systems and AI intelligent agent systems can be seamlessly connected based on a common protocol, accelerating the intelligent upgrading of industries such as smart water conservancy. At the same time, through the design of data bus, the automatic discovery and full traceability of process data are realized, which not only ensures the operation of intelligent agents, but also continuously accumulates reusable data assets.

[0034] (4) By decoupling resources and supporting data bus, cross-system general adaptation and operation optimization can be achieved, so that high stability and wide applicability can still be maintained under the condition of limited local model resources.

[0035] This invention provides a generative intelligent agent system with a long short-term memory mechanism and its implementation method, which can be widely applied to various complex and highly specialized business scenarios, such as smart water conservancy, smart water affairs, and other professional application scenarios that require customized intelligent agent workflows and autonomous intelligent agent orchestration. It also overcomes several prominent problems, as follows:

[0036] (1) To address the large model illusion problem, the introduction of long short-term memory mechanism and agent technology ensures that the large model does not go astray during operation;

[0037] (2) To address the issue of insufficient support for business scenarios by large models, the technology of converting text SOPs (Standard Operation Procedures) into workflows was introduced to achieve customized adaptation for business scenarios;

[0038] (3) To address the issue of complex workflow orchestration, a multi-agent collaborative generation technology is adopted, which allows AI to autonomously plan the working steps of the agents based on business input;

[0039] (4) To address the difficulty of maintaining intelligent agents, a rapid editing method for visual workflows based on natural language is achieved by adopting grid-based intelligent agent orchestration technology;

[0040] (5) To address the issue of the ever-increasing scale of business activities, a workflow memory mechanism is adopted to achieve mixed workflow arrangement and reference generation, thereby expanding the scope of business issues;

[0041] (5) To address the issues of continuous model optimization and miniaturization, we combine workflow memory storage with the construction of long-term working memory for knowledge retrieval or continuous model training, thereby reducing the number of model parameters while ensuring the availability of capabilities. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in this embodiment, the accompanying drawings used in the description of the embodiment will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of the agent generator generation workflow in the generative agent system with long short-term memory mechanism provided by the present invention.

[0044] Figure 2 This is a user flowchart of the generative intelligent agent system with long short-term memory mechanism provided by the present invention.

[0045] Figure 3 This is a schematic diagram of the editing page of the agent editor in the generative agent system with long short-term memory mechanism provided by the present invention;

[0046] Figure 4 This is a flowchart illustrating the application of the data bus in the generative intelligent agent system with long short-term memory mechanism provided by the present invention. Detailed Implementation

[0047] To make the technical means, creative features, and achieved objectives and effects of this invention readily understandable, the technical solution of this invention will be clearly and completely described below in conjunction with embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0048] Example 1: This example provides a generative intelligent agent system with a long short-term memory mechanism, including the following components:

[0049] The intelligent agent generator is responsible for generating workflows tailored to specific business scenarios based on user needs, textual SOPs stored in the knowledge base, and existing workflows.

[0050] The agent editor is responsible for visually displaying the workflow generated by the agent generator, where users can customize or arrange the various work steps of the agent and preview the execution effect;

[0051] The agent executor is responsible for converting the workflow text generated by the agent generator into an executable program, and for performing execution verification and feedback of execution results.

[0052] The data bus is responsible for data management within the workflow, ensuring data transmission, storage, and application during workflow execution.

[0053] The knowledge base is responsible for storing data input from external sources or data accumulated by the system during runtime. The external input data includes relevant business knowledge, terminology database, SOPs, and examples of generated workflows or SQL query statements.

[0054] The tool library is responsible for registering MCP services, discovering a list of general tools available to agents, and allowing agents to select appropriate tools to complete business functions when generating workflows.

[0055] The algorithm library, with the same execution process as the tool library, integrates independent professional business algorithms designed and developed by the business system.

[0056] Specifically, in this embodiment 1, see... Figure 1The intelligent agent generator is specifically composed of five intelligent agents: a planning intelligent agent for task decomposition, a data query intelligent agent for planning data queries, an algorithm intelligent agent for selecting business algorithms, a knowledge intelligent agent for recalling related knowledge, and a summarizing intelligent agent for summarizing workflows. These five intelligent agents generate workflows according to the multi-agent role definition and their collaboration scheme. The data query intelligent agent, algorithm intelligent agent, and knowledge intelligent agent are each responsible for the specific workflow execution content in a certain direction and report the results to the planning intelligent agent to verify the task completion status. After verification, the results are summarized.

[0057] Specifically, in this embodiment 1, see... Figure 1 The system consists of three intelligent agents: a planning agent, a data query agent, and a knowledge agent, each responsible for decomposing user requirements and distributing them to three types of agents: the data query agent, the algorithm agent, and the knowledge agent. The data query agent executes the planned data query workflow and reports the results to the planning agent to verify task completion. The algorithm agent executes the business algorithm selection workflow and reports the results to the planning agent to verify task completion. The knowledge agent executes the related knowledge retrieval workflow and reports the results to the planning agent to verify task completion. The aggregation agent aggregates the verified planned data query workflow, business algorithm selection workflow, and related knowledge retrieval workflow. In simpler terms: the planning agent decomposes the work; the data query agent, algorithm agent, and knowledge agent each select tools and execution steps based on the planned tasks to determine data requirements; the aggregation agent, based on the planning agent's tasks, divides the tools selected by the three types of agents and then coordinates the parallel and sequential execution steps according to data dependencies.

[0058] Specifically, in this embodiment 1, the method for generating workflow by the agent generator is to generate an executable workflow according to the agent generator's generation protocol by using planning agents, data lookup agents, algorithm agents, and knowledge agents.

[0059] Specifically, in this embodiment 1, the planning agent can choose to enter the direct workflow generation branch and generate all work steps itself, based on the size and difficulty of the user's problem and the existing workflows in the knowledge base, or choose to enter the multi-agent collaborative process, where multiple agents collaborate to generate all work steps. Choosing to enter the multi-agent collaborative process improves the stability of workflow generation.

[0060] Specifically, in this embodiment 1, the workflow generated by the agent generator can be further edited and published by the agent editor.

[0061] Specifically, in this embodiment 1, the agent editor can directly edit a smart agent as specified by the composite actuator protocol, or it can load a workflow generated by the agent generator and submit it to the user for editing.

[0062] Specifically, in this embodiment 1, the agent editor further edits and publishes the workflow generated by the agent generator, including the following steps:

[0063] A1. Workflow Loading: Load an existing workflow using the agent editor, and expand and edit each specific atomic step in the generated workflow;

[0064] A2. Workflow Creation: Users can manually edit the tasks to be performed and their data dependencies for each work node, and plan the execution order of each atomic node; during editing, users can directly generate a workflow template through the agent generator and make adjustments based on it;

[0065] A3. Workflow Editing: Users can adjust the specific behavior of atomic operations and their output and input data dependencies by setting prompts;

[0066] A4. Workflow Validation and Preview: The user clicks the play button at the starting node in the workflow to execute the workflow until the results of all execution steps are displayed; if there are data dependency errors or the workflow cannot run during execution, the system will report an error and explain the reason for the error.

[0067] Specifically, in this embodiment 1, the intelligent agent editor has the following key technical features:

[0068] Firstly, in the intelligent agent editor, adjacent atomic nodes belong to different execution steps in the same workflow. Work nodes in the same row can be executed in parallel, while nodes in different rows are executed serially. Data is transmitted between rows through a data bus, and each node obtains data from the data bus according to its own definition.

[0069] Secondly, by dragging and dropping atomic node blocks to different rows and columns, the execution order of the workflow can be directly adjusted. Alternatively, deleting an atomic node block can adjust the execution steps of the entire workflow. (See [link / reference]). Figure 3 ,exist Figure 3 The 5 rows and 6 columns of squares displayed in the middle are atomic node blocks.

[0070] Third, an agent editor can contain multiple workflow editor canvases, and each workflow editor canvas can contain multiple unconnected workflows. These unconnected workflows can be executed individually or synchronously via the play button on the canvas.

[0071] Fourth, a workflow editor canvas can contain multiple workflows, or a single workflow, or be published as a separate service to form a single agent service.

[0072] Specifically, in this embodiment 1, the intelligent agent executor execution protocol is defined as follows:

[0073] {

[0074] "question": "Original question description",

[0075] "reasoning": "Workflow design approach analysis",

[0076] "createDate" : "Current Time",

[0077] "userId": "user ID",

[0078] "name": "Agent Name", / / If provided above, use the data from above; otherwise, generate one according to the user's question.

[0079] "isPublic": "Whether to make public", / / 0 means not public, 1 means public

[0080] "workflow": {

[0081] "type": "sequence",

[0082] "tasks": [

[0083] {

[0084] "type": "parallel",

[0085] "taskId": "P1",

[0086] "tasks": [

[0087] / / All data acquisition tasks must be placed here. ]

[0089] },

[0090] {

[0091] "type": "atomic", / / required

[0092] "taskId": "A2", / / Required

[0093] "description": "Task description", / / Required

[0094] "tool": "Tool Name", / / Required

[0095] "parameters": {

[0096] "inputData": "{{output_of:P1}}" / / Adapt parameters according to the selected method

[0097] }

[0098] } ]

[0100] }

[0101] }

[0102] The intelligent agent executor can execute the atomic tasks and tools defined in the atomic nodes in parallel according to the above method, thereby completing the predetermined work steps of the workflow.

[0103] Specifically, in this embodiment 1, see... Figure 4 The workflow generated by the agent generator does not rely on connection-oriented deterministic data transmission. Instead, it uses a key-value format, where worker nodes specify their input data requirements and output data formats during generation. Clear data dependencies between nodes are ensured through planning and aggregating agents, and workflow validation guarantees the feasibility of these dependencies. During execution, worker node output data is stored in Redis, and the required data is transmitted to worker nodes via a data bus to ensure data availability. A worker node can be any possible form defined in the workflow, including various decomposable atomic tasks such as direct large-scale model question answering, data lookup, algorithm calculation, and business system interface calls.

[0104] Example 2: This example provides a method for implementing a generative intelligent agent system with a long short-term memory mechanism, including the following steps:

[0105] B1. Users input user questions into a generative intelligent agent system with a long short-term memory mechanism based on their needs;

[0106] B2. The generative intelligent agent system with a long short-term memory mechanism determines whether there is a workflow in its system that matches the user's question based on the user's question.

[0107] B3. If there is a workflow in the system that matches the user's problem, then the agent executor in the generative agent system with long short-term memory mechanism executes the workflow according to the workflow that matches the user's problem.

[0108] B4. If there is no workflow matching the user problem in the system, the agent generator in the generative agent system with long short-term memory mechanism first generates a workflow for the user problem. Then, the agent editor in the generative agent system with long short-term memory mechanism visualizes the workflow generated by the agent generator. After that, the agent executor in the generative agent system with long short-term memory mechanism executes the workflow generated by the agent generator. Finally, the successfully executed and published workflow generated by the agent generator for the user problem is saved in the agent memory.

[0109] Specifically, in step B4 above, the workflow generated by the agent generator and addressing the user's problem that has been successfully executed and published is stored in the agent memory. This can serve as a workflow that can be directly used by the user later, and also as an atomic tool for the agent to generate workflows.

[0110] Finally, it should be noted that the above description is only an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A generative intelligent agent system with a long short-term memory mechanism, characterized in that, include: The intelligent agent generator is responsible for generating workflows tailored to business scenarios based on user needs, textual SOPs stored in the knowledge base, and existing workflows. The intelligent agent generator consists of five intelligent agents: a planning intelligent agent for decomposing tasks, a data query intelligent agent for planning data queries, an algorithm intelligent agent for selecting business algorithms, a knowledge intelligent agent for recalling related knowledge, and a summarizing intelligent agent for summarizing workflows. The agent editor is responsible for visually displaying the workflow generated by the agent generator, where users can customize or arrange the various work steps of the agent and preview the execution effect. The workflow generated by the agent generator can be further edited and published by the agent editor. The agent editor further edits and publishes the workflow generated by the agent generator, including the following steps: A1. Workflow loading: Load an already generated workflow through the agent editor, expand and edit each specific atomic step in the generated workflow; A2. Workflow creation: The user manually edits each work step. The workflow is defined by the tasks each node needs to perform and its data dependencies, and the execution order of each atomic node is planned. During editing, users can directly generate a workflow template using the agent generator and make adjustments accordingly. A3. Workflow Editing: Users adjust the specific behaviors of atomic operations and their output and input data dependencies using prompts. A4. Workflow Validation and Preview: Users execute the workflow by clicking the play button on the starting node until the results of all execution steps are displayed. If data dependency errors occur during execution or the workflow fails to run, the system will report an error and explain the reason for the error. The agent executor is responsible for converting the workflow text generated by the agent generator into an executable program, and for performing execution verification and feedback of execution results. The data bus is responsible for data management within the workflow, ensuring data transmission, storage, and application during workflow execution. The knowledge base is responsible for storing data input from external sources or data accumulated by the system during runtime. The external input data includes relevant business knowledge, terminology database, SOPs, and examples of generated workflows or SQL query statements. The tool library is responsible for registering MCP services, discovering a list of general tools available to agents, and allowing agents to select appropriate tools to complete business functions when generating workflows. The algorithm library, with the same execution process as the tool library, integrates independent professional business algorithms for business system design and development; The agent memory is responsible for storing workflows that have been successfully executed and published, which can then be used directly by users or as atomic tools for agents to generate workflows.

2. The generative intelligent agent system with long short-term memory mechanism according to claim 1, characterized in that, The agent editor can directly edit an agent as defined by a composite actuator protocol, or it can load a workflow generated by an agent generator and submit it to the user for editing.

3. The generative intelligent agent system with long short-term memory mechanism according to claim 1, characterized in that, In the intelligent agent editor, adjacent atomic nodes belong to different execution steps in the same workflow. Work nodes in the same row can be executed in parallel, nodes in different rows are executed serially, and data is transmitted between rows through a data bus. Each node obtains data from the data bus according to its own definition.

4. The generative intelligent agent system with long short-term memory mechanism according to claim 1, characterized in that, In the intelligent agent editor, the execution order of the workflow can be directly adjusted by dragging and dropping atomic node blocks to different rows and columns. Alternatively, an atomic node can be deleted to adjust the execution steps of the entire workflow.

5. The generative intelligent agent system with long short-term memory mechanism according to claim 1, characterized in that, The intelligent agent editor contains multiple workflow editor canvases, each containing multiple unconnected workflows. These unconnected workflows can be executed individually or synchronously via the play button on the canvas.

6. The generative intelligent agent system with long short-term memory mechanism according to claim 1, characterized in that, The agent editor contains multiple workflow editor canvases. Each workflow editor canvas contains multiple workflows, or a single workflow, or it can be published as a separate service to form a separate agent service.

7. A method for implementing a generative intelligent agent system with a long short-term memory mechanism, characterized in that, Includes the following steps: B1. The user inputs a user question into the generative intelligent agent system with long short-term memory mechanism as described in any one of claims 1 to 6 according to their needs; B2. The generative intelligent agent system with long short-term memory mechanism determines whether there is a workflow in its system that matches the user's question based on the user's question. B3. If there is a workflow in the system that matches the user problem, then the agent executor in the generative agent system with long short-term memory mechanism executes the workflow according to the workflow that matches the user problem. B4. If there is no workflow in the system that matches the user problem, the agent generator in the generative agent system with long short-term memory first generates a workflow for the user problem. Then, the agent editor in the generative agent system with long short-term memory visualizes the workflow generated by the agent generator. After that, the agent executor in the generative agent system with long short-term memory executes the workflow generated by the agent generator. Finally, the successfully executed and published workflow generated by the agent generator for the user problem is saved in the agent memory.

Images