A memory bank-based multi-agent collaborative teaching content generation method and system

Abstract

This invention proposes a multi-agent collaborative teaching content generation method and system based on a memory bank, belonging to the field of intelligent teaching content generation technology. It includes: analyzing an initial decomposition plan from a global task context composed of user teaching needs using a central control agent; verifying and optimizing the initial decomposition plan through a plan verification agent and a plan optimization agent, respectively, generating a list of atomic subtasks containing dependencies and multi-agent routing signaling; constructing multimodal teaching segments using corresponding expert agents; performing quantitative verification and scoring through an evaluation agent; and outputting teaching content based on the verification results. This invention can accurately generate high-quality teaching content that matches the needs of the target users while avoiding resource waste.

Description

Technical Field

[0001] This invention belongs to the field of intelligent teaching content generation technology, and in particular relates to a method and system for generating teaching content through multi-agent collaborative interaction based on a memory bank. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of large language model technology, it has shown great application potential in areas such as assisting instruction, lesson plan generation, exercise generation, and teaching resource integration. Existing methods for generating teaching content typically rely on user-inputted teaching needs, directly generating corresponding content by calling the language model. However, when dealing with complex science education tasks with long logical chains, such as generating a physics lesson plan that simultaneously includes formula derivations, experimental procedures, and accompanying simulation animations, existing methods still have the following shortcomings: (1) Lack of dynamic decomposition and collaborative planning capabilities for complex teaching tasks. Existing teaching content generation methods often struggle to systematically break down the overall teaching task into several dependent atomic subtasks when dealing with multi-step and multi-type teaching needs. They also struggle to schedule corresponding generation capabilities according to the content type of different subtasks. When teaching content involves multiple stages such as knowledge explanation, formula derivation, code generation, animation generation, and document integration, problems such as unreasonable execution order, omission of subtasks, logical breaks in content, and difficulty in integrating generation results are prone to occur, making it difficult for the final generated teaching materials to directly meet the usage needs of complex teaching scenarios.

[0004] (2) Lack of a systematic accumulation and reuse mechanism for historical task experience. When dealing with new teaching tasks, existing content generation methods usually rely mainly on the model's own pre-trained knowledge and current input information, making it difficult to structure and reuse successful generation paths, effective prompt samples, code examples, instructional design strategies, and failure reflection information from historical tasks. Therefore, when faced with teaching tasks with similar structures or similar knowledge points, the system often needs to repeat the same or similar reasoning, generation, and debugging processes, which not only increases the consumption of computing resources but also limits the system's ability to continuously optimize as the number of uses increases. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a method and system for generating multi-agent collaborative teaching content based on a memory bank, which can accurately generate high-quality teaching content that matches the needs of target users while avoiding resource waste.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for generating multi-agent collaborative teaching content based on a memory bank.

[0007] A method for generating multi-agent collaborative teaching content based on a memory bank includes: The target user's historical learning characteristics and current teaching needs are used as the global task context; The initial decomposition plan is obtained by encoding and similarity matching of the global task context by the overall control agent; the initial decomposition plan is verified and optimized by the plan verification agent and the plan optimization agent respectively, generating a list of atomic subtasks containing dependencies and multi-agent routing signaling; The expert agents in the expert agent cluster, corresponding to the atomic subtask list and multi-agent routing signaling, obtain sample prompts and scientific fact constraints from the memory bank to construct multimodal teaching segments. The multimodal teaching segments are quantitatively verified and scored by the evaluation agent, and the verified multimodal teaching segments are output as the final teaching content.

[0008] Furthermore, the memory bank includes a planning memory bank, an execution memory bank, a scientific knowledge memory bank, and a user preference memory bank.

[0009] Furthermore, the expert agent cluster consists of multiple expert agents, including a knowledge generation agent, a vision generation agent, an instructional design agent, and a document management agent; each expert agent executes atomic subtasks based on the corresponding multi-agent routing signaling and calls internal tools.

[0010] Furthermore, the evaluation agent performs quantitative verification and scoring of multimodal teaching segments, including: if the score is lower than a preset threshold, criticism features are generated and returned to the multi-agent execution module and task orchestration and decomposition module for iterative refinement; if the score meets the standard, the verified multimodal teaching segments are output as the final teaching content that meets the teaching needs of the target users.

[0011] Furthermore, the evaluation agent performs quantitative verification and scoring on multimodal teaching segments through sandbox verification, semantic consistency comparison, and multi-dimensional quantitative evaluation.

[0012] Furthermore, a multi-agent collaborative teaching content generation method based on a memory bank also includes a memory update process. Specifically, the target user's preference path and reflections on failure cases are synchronously written into the corresponding memory bank to achieve incremental updates of the memory bank.

[0013] A second aspect of the present invention provides a multi-agent collaborative teaching content generation system based on a memory bank.

[0014] A multi-agent collaborative teaching content generation system based on a memory bank includes: a data storage layer, a collaborative scheduling layer, and an evaluation and self-learning layer; The data storage layer consists of multiple memory banks, specifically including a planning memory bank, an execution memory bank, a scientific knowledge memory bank, and a user preference memory bank; The collaborative scheduling layer is constructed based on a state-driven directed graph framework, specifically including a central control agent, a plan verification agent, a plan optimization agent, and an expert agent cluster composed of multiple expert agents; The evaluation and self-learning layer consists of an evaluation agent module, which is used to perform multi-dimensional verification of multimodal teaching segments.

[0015] Furthermore, the data storage layer, collaborative scheduling layer, and evaluation and self-learning layer are specifically implemented through the following core modules: The dynamic memory and update module are configured as follows: as a data storage layer, it integrates a planning memory, an execution memory, a scientific knowledge memory, and a user preference memory; and it synchronously writes the target user's preference path and reflections on failure cases into the corresponding memory to achieve incremental updates of the memory. The user needs analysis module is configured to: obtain the teaching needs information of the target user, extract the core teaching needs features from the teaching needs information, and use the target user's historical learning features and current teaching needs features as the global task context to send them to the task orchestration and decomposition module. The task orchestration and decomposition module is configured to: encode and match the global task context based on the overall control agent to obtain an initial decomposition plan; verify and optimize the initial decomposition plan through the plan verification agent and the plan optimization agent respectively, generate an atomic subtask list containing dependencies and multi-agent routing signaling, and send it to the multi-agent execution module; The multi-agent execution module is configured to: employ expert agents in the expert agent cluster that correspond to the atomic subtask list and multi-agent routing signaling, retrieve sample prompts and scientific fact constraints from the memory bank, construct multimodal teaching segments, and distribute them to the content evaluation and optimization module; The content evaluation and optimization module is configured to: use a built-in evaluation agent to quantitatively verify and score the multimodal teaching segments, and output the verified multimodal teaching segments as the final teaching content.

[0016] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a memory-based multi-agent collaborative teaching content generation method as described in the first aspect of the present invention.

[0017] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a memory-based multi-agent collaborative teaching content generation method as described in the first aspect of the present invention.

[0018] The above one or more technical solutions have the following beneficial effects: (1) When faced with complex tasks that require combining multiple stages such as knowledge explanation, code generation, and content integration, the overall control agent can dynamically decompose the global task context into a list of atomic subtasks containing dependencies and routing signals; the plan verification agent evaluates the completeness, executability, and dependency rationality of the decomposed plan from multiple dimensions; and the plan optimization agent iteratively adjusts the plan based on the evaluation feedback. This mechanism effectively overcomes the shortcomings of existing technologies in forming structured execution plans, ensures the logic and executability of task decomposition, fundamentally avoids the chaos in the generation process and the break in content logic, and can stably produce high-quality teaching materials that can directly meet the needs of complex teaching scenarios.

[0019] (2) This invention establishes a systematic experience storage and reuse channel by constructing a data storage layer that includes a planning memory and an execution memory, and combining a dynamic memory and an update module. When processing teaching tasks with similar structures, the overall control agent can retrieve successful or failed task decomposition cases from the planning memory as a reference, and the expert agent can retrieve historical successful generation trajectories from the execution memory as a few-sample prompt. More importantly, after a task is successfully completed or after a failure is reflected upon, the system will synchronously write the task decomposition path, execution trajectory, and reflection information into the corresponding memory, realizing incremental updates of memory. This mechanism enables this invention to accumulate experience from historical tasks, effectively overcoming the shortcomings of existing technologies that require reasoning from scratch for each task and cannot reuse effective paths, significantly reducing the overhead of repeated calculations, and enabling the system performance to continuously optimize and iterate with the increase in the number of uses.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of a multi-agent collaborative teaching content generation method based on a memory bank, according to Embodiment 1 of the present invention.

[0023] Figure 2 This is a flowchart of task decomposition and optimization in Embodiment 1 of the present invention.

[0024] Figure 3 This is a schematic diagram of the multi-dimensional criticism feedback and decomposition optimization closed loop in Embodiment 1 of the present invention.

[0025] Figure 4 This is an architectural flowchart of a multi-agent collaborative teaching content generation system based on a memory bank, as described in Embodiment 2 of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Example 1 This embodiment discloses a method for generating multi-agent collaborative teaching content based on a memory bank.

[0030] like Figure 1 As shown, a method for generating multi-agent collaborative teaching content based on a memory bank includes: Step S1: Use the target user's historical learning characteristics and current teaching needs as the global task context; Step S2: Based on the overall control agent, the global task context is encoded and similarity is matched to obtain the initial decomposition plan; the initial decomposition plan is verified and optimized by the plan verification agent and the plan optimization agent respectively, generating a list of atomic subtasks containing dependencies and multi-agent routing signaling; Step S3: Using the expert agents in the expert agent cluster that correspond to the atomic subtask list and multi-agent routing signaling, sample prompts and scientific fact constraints are obtained from the memory bank to construct multimodal teaching segments; Step S4: The multimodal teaching segments are quantitatively verified and scored by the evaluation agent, and the verified multimodal teaching segments are output as the final teaching content.

[0031] Based on the above process, this invention can accurately generate high-quality teaching content that matches the needs of target users while avoiding resource waste. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0032] In step S1, the target user's historical learning characteristics and current teaching needs characteristics are used as the global task context.

[0033] By acquiring teaching needs information input by target users, a lightweight large language model combined with a domain knowledge graph is used for semantic parsing to extract core needs features; and a planning memory is established. Execution memory Scientific knowledge memory bank and user preference memory Extract the target user's historical learning features and combine the core demand features with the historical learning features to generate a global task context. .

[0034] The planning memory, execution memory, scientific knowledge memory, and user preference memory together constitute the data storage layer, specifically: Planning memory bank Used to store original task and subtask plans. (i.e., the sequence of subtasks after task decomposition), planning status (the quality assessment results of the planning include success and failure), failure cases and their reflection information (including criticism features). The planning memory uses a vector database for storage, uses task descriptions and subtask planning as retrieval units, and supports filtering according to planning status.

[0035] Execution Memory This system primarily stores the trajectory and knowledge of successful task execution. The stored content includes task description, execution trajectory, execution status, posterior knowledge, task hierarchy, and task identifier. The storage structure mainly employs a vector database, using an embedding model to convert task descriptions into vector representations. It supports metadata filtering and allows retrieval based on conditions such as execution status and task hierarchy. The retrieval mechanism is based on vector similarity. Retrieve similar tasks, support filtering by execution status (success, failure), and return the Top-K most similar tasks and their execution trajectories: ; in, This indicates the currently pending task or the current atomic subtask. This represents the historical tasks stored in the memory bank; This represents an embedding encoding function used to map a task description, task context, or task execution trajectory into a vector representation. This indicates the similarity between the current task and historical tasks.

[0036] Based on this, the system follows The historical tasks are sorted from highest to lowest value, and the tasks with the highest similarity are selected. A historical task and its corresponding execution trajectory serve as a reference sample for the current task.

[0037] Scientific knowledge memory bank This knowledge memory primarily stores domain-specific concepts, formulas, and principles, as well as core objective knowledge from scientific fields such as physics, chemistry, and biology. It extracts key concepts and knowledge from task descriptions and execution processes, establishes relationships between these concepts, forms a knowledge graph, and retrieves relevant domain knowledge based on task requirements.

[0038] User preference memory It stores users' learning stages, content preferences, historical learning records, preference path patterns, etc., using key-value pairs for storage.

[0039] In step S2, the global task context is encoded and similarity matched based on the overall control agent to obtain the initial decomposition plan; the initial decomposition plan is verified and optimized by the plan verification agent and the plan optimization agent respectively to generate a list of atomic subtasks containing dependencies and multi-agent routing signaling.

[0040] like Figure 2 As shown, the global task context is received through the central control agent. Encode its features Then, combined with the planning memory bank Similarity matching of successful and unsuccessful cases. Dynamically generate initial decomposition plan In the specific implementation process, the first step is to plan the memory bank. Search for similar historical cases in the middle and calculate If the search results are found The matching degree is higher than the preset matching threshold If a successful case template is available, refer to that template to generate an initial breakdown plan; if If the matching degree is lower than the matching threshold or a failed case is retrieved, the dynamic generation process is initiated, that is: the central control agent combines the scientific knowledge memory base. Using relevant domain knowledge and the reasoning capabilities of a large language model, a list of atomic subtasks containing dependencies is dynamically generated. And incorporate user preference memory into the generation process. The system stores users' historical preference paths (e.g., users previously preferred "concept explanations followed by animated demonstrations"), ensuring that the breakdown plan aligns with users' personalized learning habits. Among these, the matching threshold... In this embodiment, the value is 0.8.

[0041] Subsequently, through collaboration with the plan verification agent and the plan optimization agent, the initial decomposition plan is optimized. After verification and iterative optimization, atomic subtasks containing dependencies are finally generated. List and multi-agent routing signaling Specifically, a plan verification agent is introduced to validate the initial decomposition plan. A multi-dimensional evaluation is conducted, including: completeness (whether it covers all teaching points), executability (whether each atomic subtask has corresponding expert agent support), dependency rationality (whether the dependencies between subtasks conform to the teaching logic), and user preference matching (whether it fits the user's historical learning path). The validation results are output as a score or pass / fail. If the plan validation fails, a plan optimization agent is activated to generate optimization suggestions based on the validation feedback, adjusting the decomposed plan (e.g., splitting overly coarse subtasks, adjusting the execution order, supplementing missing steps, etc.), iteratively optimizing until validation is passed or the maximum number of iterations (e.g., 3 times) is reached. The optimized final plan serves as the basis for task execution.

[0042] Among them, the overall control agent is responsible for the global task context. Processing, dynamic task decomposition (combined with user preference paths), and generation of initial decomposition plans. It also coordinates the plan verification and optimization process. The plan verification agent is responsible for decomposing the plan. The system evaluates completeness, executability, dependency rationality, and user preference fit, and outputs the verification results. The planning optimization agent is responsible for generating optimization suggestions based on verification feedback and adjusting the decomposed plan when verification fails. Until it passes verification.

[0043] In step S3, the expert agents in the expert agent cluster that correspond to the atomic subtask list and multi-agent routing signaling are used to obtain sample prompts and scientific fact constraints from the memory bank to construct multimodal teaching segments.

[0044] With multiple built-in atomic tasks A cluster of expert agents with matched types receives atomic subtasks. and routing signaling This will activate the corresponding expert agent. The expert agent retrieves data from its execution memory. Retrieving Top-K few-sample hints from the scientific knowledge memory database The system obtains scientific fact constraints; based on the retrieved memory information and scientific fact constraints, the expert agents work collaboratively according to their dependencies to generate multimodal teaching segments such as text, images, videos, and code.

[0045] The expert agent cluster includes knowledge-generating agents, vision-generating agents, instructional design agents, and document management agents. Each expert agent specifically implements functions for its respective domain, based on routing signaling. Execute the corresponding atomic subtasks Each component utilizes all its tools and receives feedback to refine and optimize its content.

[0046] Among them, the central control agent and the expert agent share a state space to achieve efficient collaboration in task distribution, result feedback and data interaction.

[0047] In step S4, the multimodal teaching segments are quantitatively verified and scored by the evaluation agent, and the verified multimodal teaching segments are output as the final teaching content.

[0048] like Figure 3 As shown, the intelligent agent is evaluated and scored quantitatively across three dimensions: consistency with scientific facts, rationality of educational logic, and multimodal alignment. If the score is below a preset threshold, criticism features are generated and returned to the multi-agent execution module and task orchestration and decomposition module for iterative refinement. If the score meets the standard, the validated multimodal teaching segments are sent to the dynamic memory and update module, and the final teaching content is output to the user. The quantitative verification threshold can be set according to actual needs; this embodiment does not impose a specific limitation on it.

[0049] Furthermore, in order to quantitatively verify and score multimodal teaching segments from three dimensions—consistency of scientific facts, rationality of educational logic, and multimodal alignment—this invention uses sandbox verification, semantic consistency comparison, and multi-dimensional quantitative evaluation to quantitatively verify and score multimodal teaching segments, specifically as follows: ; in, This represents the overall quality score of a multimodal teaching segment; Indicates the first Each evaluation indicator has a corresponding weight coefficient, and the sum of the weights of each evaluation indicator is 1. Indicates the first The score value corresponding to each evaluation indicator. This indicates the number of evaluation indicators involved in the quantitative verification. When the evaluation dimensions include three dimensions: consistency with scientific facts, rationality of educational logic, and multimodal alignment, this is relevant. .

[0050] Based on this, critical features containing decomposition suggestions are generated, triggering a double closed loop of content refinement and decomposition plan adjustment.

[0051] Once the multimodal teaching segment passes verification and is successfully output, a synchronous experience storage mechanism is triggered: The successfully executed task breakdown path and user-preferred execution path (such as user-preferred sub-task sequences) are vectorized and appended to the planning memory according to similarity deduplication rules. The system stores the failed cases and their detailed reflections (including reasons for failure, critical features, optimization attempts, etc.) in the failure case section of the planning memory after structuring them. The successfully generated problem-solving steps, multimodal rendering code, and generation trajectory are also structured and appended to the execution memory. Finally, user preferences reflected in this task (such as positive user feedback on a certain type of teaching method) are updated to the user preference memory, achieving incremental self-updating of the memory. The specific incremental update formula is as follows: ; in, This indicates the updated memory. This refers to the memory before the update. This represents the incremental memory data generated during the execution of this task.

[0052] To further facilitate understanding of the technical solution of this invention, this embodiment uses the user's request "Help me generate a lesson plan for high school projectile motion, along with a simulation animation of the motion trajectory" as an example to detail the entire execution steps of a multi-agent collaborative teaching content generation method based on a memory bank. This method embodies core improvements such as dynamic task decomposition, user preference fusion, plan verification and optimization, multi-dimensional feedback (including decomposition suggestions), and synchronous memory updates. The implementation of the method specifically includes the following steps: The first step is to parse the user request and extract the preceding state.

[0053] After receiving the user's input request, the system calls the natural language processing model to perform semantic parsing of the request. The core request features extracted are shown in Table 1.

[0054] Table 1 Examples of Core Requirements Features

[0055] At the same time, from the user preference memory bank The historical learning characteristics of users were retrieved, and the retrieved historical learning characteristics are shown in Table 2.

[0056] Table 2 Examples of Historical Learning Characteristics

[0057] The core requirements features are combined with historical learning features to generate a global task context. This information is then passed to the task orchestration and decomposition module.

[0058] The second step is task decomposition and optimization based on dynamic programming memory.

[0059] 1) Dynamic task decomposition.

[0060] The central control agent receives the global task context. And encode its vector features, first in the planning memory. Similarity retrieval is performed, and calculation is performed. Search for historical high-confidence task breakdown templates; the search revealed a successful template match rate of 0.82 related to "generating high school physics kinematics knowledge points". (Higher than the preset threshold), but the template does not include the user's preferred animation demonstration path. At the same time, the planning memory also stores failure cases: once, the user requested to add it again because of the omission of "the lack of visualization after the formula derivation". The criticism features recorded the suggestion of "adding a principle visualization subtask".

[0061] The central control agent dynamically generates an initial decomposition plan by combining retrieved successful templates, reflections on failed cases, and preference paths ("concept explanation" → "animation demonstration" → "experimental verification") from the user preference memory. The plan includes the following atomic subtasks and dependencies to generate routing signaling. Specify the executing agent: Subtask A: Extract the core physical formulas of projectile motion, explain the core concepts and formulas of projectile motion, and route it to a knowledge generation agent; this is text generation and has no dependencies. Subtask B: Generate simulation code for projectile motion trajectories based on physical formulas, routed to a vision-generated intelligent agent; this is code generation and depends on subtask A (requires formula input). Subtask C: Design a lesson plan document containing experimental steps, and extract 4 keyframes evenly from the animation generated by subtask B as illustrations for the motion trajectory part of the lesson plan. This is routed to the document management agent, belongs to document generation, and depends on subtask A and subtask B. Subtask D: Integrate animated demonstrations and lesson plan documents to generate the final teaching package, which is then routed to the file management agent. This involves file packaging and depends on subtask C.

[0062] The breakdown plan shows a path that incorporates user preferences for "animations following explanations," placing animation generation after formula explanations and adding keyframes of the generated simulation animation as illustrations in the lesson plan, in sections that match the animation content.

[0063] 2) Plan verification.

[0064] The plan is to verify the agent's understanding of the initial decomposition plan. A multi-dimensional assessment was conducted as shown in Table 3.

[0065] Table 3 Examples of historical learning characteristics.

[0066]

[0067] If the validation fails (for example, the user preference path is not adopted), the planning optimization agent will generate optimization suggestions based on the validation feedback, adjust the decomposition plan, and continue until the validation is passed.

[0068] The third step is to implement the coordinated execution of executive memory and knowledge memory.

[0069] Based on the optimized task plan, the central control agent generates multi-agent routing signaling to wake up the corresponding expert agents. Before execution, each agent retrieves relevant information from the execution memory and scientific knowledge memory. Knowledge-generating intelligent agent (performing subtask A): from scientific knowledge memory The core constraint formula for projectile motion was found in the search results: , ;in, It represents the horizontal displacement of an object. This represents the initial horizontal velocity of the object when it is thrown. It represents the vertical displacement of an object. This represents the time of motion of the object measured from the moment it is launched. Simultaneously, a previously successfully generated "free fall" explanation text template is retrieved from the execution memory as a few-sample prompt to generate formula explanation text, noting that an animated demonstration is coming soon. The generated formula parsing text is then passed to the visual generation agent and the file management agent.

[0070] Visual generative agent (performing subtask B): First, from the scientific knowledge memory... The scientific representation constraints of the projectile motion trajectory are obtained (the trajectory should be a parabola); subsequently, the execution experience base is used. The system retrieved the historically successfully generated "free fall motion" simulation code, extracted it directly as a few-sample hint, and rewrote it to quickly generate simulation code for projectile motion trajectory. The code has reserved parameter interfaces for use in teaching materials, which greatly improves the first compilation success rate of the code.

[0071] The file management agent (executes subtasks C and D): from the execution memory. The system retrieves high school physics lesson plan format specifications from the database, integrates formula analysis text, simulation code, and experimental steps (obtained from a scientific knowledge memory bank), generates a complete lesson plan document, embeds drawing code into the document, and finally packages it into a teaching resource package.

[0072] The fourth step is to perform multi-dimensional internal consistency verification and refine the closed loop.

[0073] After generating the initial lesson plan and simulation code, the evaluation agent performs multi-dimensional quantitative verification, with a preset quality threshold. (0-1 scoring system), the dimensions of verification include: 1) Code Sandbox Testing: Run the projectile motion simulation code in a safe sandbox to check for code errors and whether the motion trajectory rendering conforms to the physical formula constraints; this run passed. ; 1 point.

[0074] 2) Scientific Fact Verification: Comparing the formulas and experimental procedures in the lesson plan with the objective laws in the scientific knowledge memory bank, it was found that the formulas were correct, but the experimental procedures omitted the crucial step of "measuring horizontal displacement," which did not conform to the standard procedures in the memory bank. The scientific factual consistency score is 0.8.

[0075] 3) Logical Reasonableness of Teaching: Evaluate whether the lesson plan structure meets the requirements of first-year high school physics teaching and whether the content difficulty is appropriate for the target learning stage. The lesson plan, from concept introduction to formula derivation to animation demonstration, has a clear logic. Score: 0.9

[0076] 4) Multimodal alignment: Check whether the formulas and experimental steps in the lesson plan are semantically consistent with the motion trajectory in the simulation animation (e.g., "the ball performs projectile motion and the trajectory is a parabola" matches the animation). Score: 0.8; Overall score Represented as: ; in, , , and These represent the scores for code sandbox testing, scientific fact verification, pedagogical logical rationality, and multimodal alignment, respectively. , , and These represent the weights corresponding to code sandbox testing, scientific fact verification, pedagogical logical rationality, and multimodal alignment, respectively.

[0077] Score for this game: The score was above the preset threshold of 0.85, but the scientific fact items were too low. The evaluation of the agent's generated criticism features indicated that "the 'measuring horizontal displacement' step is missing from the experimental procedure. It is recommended to add this experimental step to subtask C and confirm its independence in the decomposition plan." This criticism feature, on the one hand, retreats along the directed graph path to the file management agent for content refinement, and on the other hand, feeds back to the overall control agent and the plan optimization agent, suggesting that the "experimental procedure design" of subtask C in the current decomposition plan may not be detailed enough. Future similar tasks should consider breaking down the experimental steps into more detailed atomic subtasks. Based on the criticism features, the file management agent retrieves information from the scientific knowledge memory. The experimental steps were completed and the lesson plan was updated. The optimized lesson plan was resubmitted for evaluation, and all aspects met the standards.

[0078] If the sandbox test reports an error or the overall score is below 0.8, the evaluation agent will generate a critical feature containing error logs and back to the corresponding execution agent via a directed graph path for modification until the verification result reaches the preset threshold.

[0079] The fifth step involves high-confidence experience synchronization storage and self-updating of the memory bank.

[0080] Once the lesson plan and simulation animation for projectile motion are successfully displayed to the user and receive positive feedback such as "acceptance" and "likes", the system immediately triggers the synchronous experience storage mechanism.

[0081] After vectorizing the task breakdown path (subtasks A→B→C→D, and marking the user preference path "concept→animation→lesson plan") and "extract formula→code generation→document summary", successful case partitions are synchronously added to the planning memory bank according to the similarity deduplication rule.

[0082] The Python simulation code for projectile motion, the lesson plan format specifications, experimental design examples, and tips for common errors in code rewriting (e.g., pay attention to the value of gravitational acceleration g) that were successfully run this time were structured and synchronously written into the execution memory. middle.

[0083] Update the user's preference for the animation demonstration (the part of the animation the user liked again in this task) to the user preference memory. In addition, strengthening the weight of this preference can provide support for the generation of personalized teaching content in the future.

[0084] The "omission of experimental steps" issues that occurred during the execution process, along with their critical characteristics and optimization processes, will be treated as failure case reflection information. This information will be structured and synchronously written into the failure case section of the planning memory, and linked with successful cases for subsequent retrieval and alerts.

[0085] This embodiment achieves the following through dynamic decomposition (integrating user preference paths), plan verification and optimization, feedback-guided decomposition adjustments, and synchronous memory updates: 1) Personalized adaptation: the decomposition plan aligns with user preferences, resulting in more user-approved content; 2) Quality assurance: the plan verification part ensures the rationality of the decomposition from the source, and feedback-guided closed-loop optimization of content and decomposition; 3) System self-updating: the memory bank continuously enriches its content, and the system's intelligence level is iteratively improved.

[0086] Example 2 This embodiment discloses a multi-agent collaborative teaching content generation system based on a memory bank.

[0087] A multi-agent collaborative teaching content generation system based on a memory bank adopts a layered architecture and modular design, including: a data storage layer, a collaborative scheduling layer, and an evaluation and self-learning layer; The data storage layer consists of multiple memory banks, specifically including a planning memory bank, an execution memory bank, a scientific knowledge memory bank, and a user preference memory bank; The collaborative scheduling layer is constructed based on a state-driven directed graph framework, specifically including a central control agent, a plan verification agent, a plan optimization agent, and an expert agent cluster composed of multiple expert agents; The evaluation and self-learning layer consists of an evaluation agent module, which is used to perform multi-dimensional verification of multimodal teaching segments.

[0088] like Figure 4 As shown, the data storage layer, collaborative scheduling layer, and evaluation and self-learning layer are implemented through the following core modules: The dynamic memory and update module are configured as follows: as a data storage layer, it integrates a planning memory, an execution memory, a scientific knowledge memory, and a user preference memory; and it synchronously writes the target user's preference path and reflections on failure cases into the corresponding memory to achieve incremental updates of the memory. The user needs analysis module is configured to: obtain the teaching needs information of the target user, extract the core teaching needs features from the teaching needs information, and use the target user's historical learning features and current teaching needs features as the global task context to send them to the task orchestration and decomposition module. The task orchestration and decomposition module is configured to: encode and match the global task context based on the overall control agent to obtain an initial decomposition plan; verify and optimize the initial decomposition plan through the plan verification agent and the plan optimization agent respectively, generate an atomic subtask list containing dependencies and multi-agent routing signaling, and send it to the multi-agent execution module; The multi-agent execution module is configured to: employ expert agents in the expert agent cluster that correspond to the atomic subtask list and multi-agent routing signaling, retrieve sample prompts and scientific fact constraints from the memory bank, construct multimodal teaching segments, and distribute them to the content evaluation and optimization module; The content evaluation and optimization module is configured to: use a built-in evaluation agent to quantitatively verify and score the multimodal teaching segments, and output the verified multimodal teaching segments as the final teaching content.

[0089] Furthermore, the dynamic memory bank and update module provides memory data retrieval services for other modules; it receives verified multimodal teaching segments sent by the content evaluation and optimization module, and appends successful experiences (including user preference paths), failure cases and reflection information to the corresponding memory bank after vector featureization and structuring through a synchronous storage mechanism, and updates the user learning data in the user preference memory bank in real time.

[0090] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0091] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a memory-based multi-agent collaborative teaching content generation method as described in Embodiment 1 of this disclosure.

[0092] Example 4 The purpose of this embodiment is to provide an electronic device.

[0093] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in a memory-based multi-agent collaborative teaching content generation method as described in Embodiment 1 of this disclosure.

[0094] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0095] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0096] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for generating multi-agent collaborative teaching content based on a memory bank, characterized in that, include: The target user's historical learning characteristics and current teaching needs are used as the global task context; An initial decomposition plan is obtained by encoding the global task context and performing similarity matching based on the overall control agent; The initial decomposition plan is verified and optimized by a plan verification agent and a plan optimization agent, respectively, generating a list of atomic subtasks containing dependencies and multi-agent routing signaling; The expert agents in the expert agent cluster, corresponding to the atomic subtask list and multi-agent routing signaling, obtain sample prompts and scientific fact constraints from the memory bank to construct multimodal teaching segments. The multimodal teaching segments are quantitatively verified and scored by the evaluation agent, and the verified multimodal teaching segments are output as the final teaching content.

2. The method for generating multi-agent collaborative teaching content based on a memory bank as described in claim 1, characterized in that, The memory bank includes a planning memory bank, an execution memory bank, a scientific knowledge memory bank, and a user preference memory bank.

3. The method for generating multi-agent collaborative teaching content based on a memory bank as described in claim 1, characterized in that, The expert agent cluster consists of multiple expert agents, including a knowledge generation agent, a vision generation agent, an instructional design agent, and a document management agent; each expert agent executes atomic subtasks based on the corresponding multi-agent routing signaling and calls internal tools.

4. The method for generating multi-agent collaborative teaching content based on a memory bank as described in claim 1, characterized in that, The evaluation agent performs quantitative verification and scoring of multimodal teaching segments, including: if the score is lower than a preset threshold, criticism features are generated and returned to the multi-agent execution module and task orchestration and decomposition module for iterative refinement; if the score meets the standard, the verified multimodal teaching segments are output as the final teaching content that meets the teaching needs of the target users.

5. The method for generating multi-agent collaborative teaching content based on a memory bank as described in claim 4, characterized in that, The evaluation agent performs quantitative verification and scoring on multimodal teaching segments through sandbox verification, semantic consistency comparison, and multi-dimensional quantitative evaluation.

6. The method for generating multi-agent collaborative teaching content based on a memory bank as described in claim 1, characterized in that, It also includes a memory update process, specifically, writing the target user's preference path and reflections on failure cases into the corresponding memory bank in order to achieve incremental updates of the memory bank.

7. A multi-agent collaborative teaching content generation system based on a memory bank, characterized in that, include: Data storage layer, collaborative scheduling layer, and evaluation and self-learning layer; The data storage layer consists of multiple memory banks, specifically including a planning memory bank, an execution memory bank, a scientific knowledge memory bank, and a user preference memory bank; The collaborative scheduling layer is constructed based on a state-driven directed graph framework, specifically including a central control agent, a plan verification agent, a plan optimization agent, and an expert agent cluster composed of multiple expert agents; The evaluation and self-learning layer consists of an evaluation agent module, which is used to perform multi-dimensional verification of multimodal teaching segments.

8. The multi-agent collaborative teaching content generation system based on a memory bank as described in claim 7, characterized in that, The data storage layer, collaborative scheduling layer, and evaluation and self-learning layer are implemented through the following core modules: The dynamic memory and update module are configured as follows: as a data storage layer, it integrates a planning memory, an execution memory, a scientific knowledge memory, and a user preference memory; and it synchronously writes the target user's preference path and reflections on failure cases into the corresponding memory to achieve incremental updates of the memory. The user needs analysis module is configured to: obtain the teaching needs information of the target user, and extract the core teaching needs features from the teaching needs information; The task context is then distributed to the task orchestration and decomposition module, using the target user’s historical learning characteristics and current teaching needs as the global task context. The task orchestration and decomposition module is configured to: encode the global task context and perform similarity matching based on the overall control agent to obtain an initial decomposition plan; The initial decomposition plan is verified and optimized by the plan verification agent and the plan optimization agent, respectively, generating a list of atomic subtasks containing dependencies and multi-agent routing signaling, which is then sent to the multi-agent execution module. The multi-agent execution module is configured to: employ expert agents in the expert agent cluster that correspond to the atomic subtask list and multi-agent routing signaling, retrieve sample prompts and scientific fact constraints from the memory bank, construct multimodal teaching segments, and distribute them to the content evaluation and optimization module; The content evaluation and optimization module is configured to: use a built-in evaluation agent to quantitatively verify and score the multimodal teaching segments, and output the verified multimodal teaching segments as the final teaching content.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the multi-agent collaborative teaching content generation method based on a memory bank as described in any one of claims 1-6.

10. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the multi-agent collaborative teaching content generation method based on a memory bank as described in any one of claims 1-6.

Images