Memory enhancement and multi-agent collaborative generative engine optimization method and system

Abstract

The invention discloses a memory enhancement and multi-agent collaborative generative engine optimization method and system. The method comprises the following steps: obtaining a user query and initial retrieval document list; constructing a reference branch and an optimization branch; constructing a virtual editing team comprising a preference agent, a planning agent, an editing agent and an evaluation agent; the preference agent constructs a preference portrait of the specific generative engine based on large-scale historical data; executing multiple rounds of evolution type optimization iteration in the optimization branch by using the virtual editing team to generate a final optimization document; the planning agent generates an optimization instruction in combination with a preference portrait and a double-layer memory mechanism, the editing agent generates a plurality of candidate variants according to the optimization instruction, the evaluation agent screens and scores the candidate variants based on a biaxial semantic measurement standard, and the optimal variant is selected as a reference of next round of optimization; executing closed-loop memory solidification; and abstracting the optimization track of the final optimization document into a natural language rule and storing the natural language rule into a creator-level memory to update a double-layer memory mechanism.

Description

Technical Field

[0001] This invention relates to the fields of Internet information processing and artificial intelligence technology, and in particular to a generative engine optimization (GEO) method and system that combines memory enhancement and multi-agent collaboration. Background Technology

[0002] With the rapid development of Large Language Model (LLM) technology, traditional link-ranked search engines are evolving into Generative Engines (GEs) based on Retrieval Augmentation (RAG). Generative engines can directly synthesize multiple retrieved documents to generate natural language responses with citations. This paradigm shift presents content creators with a new challenge: how to achieve higher visibility and semantic influence in the responses from generative engines.

[0003] Existing optimization methods (traditional SEO or early GEO exploration) mostly rely on rule-based heuristic modifications, such as simply increasing keyword density, adding explicit citations, or modifying article structure. These methods typically treat the optimization process as a one-off text editing task for a specific query.

[0004] However, this method has obvious limitations in practical applications:

[0005] 1. Black-box unpredictability: The internal retrieval and generation logic of generative engines is an opaque "black box." Existing methods cannot distinguish between the improvements brought about by content optimization and the interference caused by natural fluctuations in search ranking, making it difficult to scientifically quantify the true causal effect of optimization measures.

[0006] 2. The evaluation metrics are one-sided and misleading: Existing evaluation systems often only focus on "apparent visibility" (such as the number of times it is cited) while ignoring "content authenticity." This makes optimization models prone to falling into an "adversarial attack" mode, using keyword stuffing or fake citations to deceive users into high exposure, creating a false sense of reality and seriously damaging the authenticity of the content.

[0007] 3. Catastrophic forgetting of strategies: Existing single-step optimization methods lack a memory mechanism. When optimizing for different queries, the model cannot accumulate historical experience, resulting in the need to start from scratch when facing similar scenarios, and it cannot adapt to the differentiated preferences of different engines.

[0008] Due to the aforementioned problems, the field of generative engine optimization urgently needs an innovative technical solution that can explicitly decouple retrieval and generation interference, balance high exposure and high fidelity, and possess continuous learning capabilities. Summary of the Invention

[0009] To address the problems in existing technologies, such as the inability to accurately quantify optimization effects due to black-box interference and content distortion caused by solely pursuing exposure, this invention provides a generative engine optimization method and system that combines memory enhancement and multi-agent collaboration. By constructing a two-branch controlled experimental protocol to freeze the retrieval context, utilizing a multi-agent co-evolutionary mechanism to optimize content in a controlled environment, and introducing a dual-layer memory mechanism to achieve lifelong learning of the strategy, this significantly improves the overall influence and attribution accuracy of content in generative responses.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] The first aspect is the generative engine optimization method that combines memory enhancement and multi-agent collaboration, which includes the following steps:

[0012] S1. Obtain the user query and initial retrieval document list of the target generative engine; construct a two-branch control experimental environment including a baseline branch and an optimization branch;

[0013] S2. Construct and initialize a virtual editing team that includes a preference agent, a planning agent, an editing agent, and an evaluation agent; the preference agent constructs a preference profile for a specific generative engine based on large-scale historical data;

[0014] S3. Use a virtual editing team to perform multiple rounds of evolutionary optimization iterations in the optimization branch to generate the final optimized document. In each iteration, the planning agent combines preference profiles and a two-layer memory mechanism to generate optimization instructions, the editing agent generates multiple candidate variants according to the optimization instructions, and the evaluation agent screens and scores the candidate variants based on a two-axis semantic metric, and selects the best variant as the benchmark for the next round of optimization.

[0015] S4. After the optimization task is completed, execute closed-loop memory solidification; abstract the final optimization document optimization trajectory into natural language rules and store it in creator-level memory to update the dual-layer memory mechanism.

[0016] As a preferred option, in S1, the construction of the two-branch control experimental environment includes:

[0017] Establish a baseline branch, keeping the original order and content of the initial searched document list unchanged, and input the target generative engine to generate a baseline response;

[0018] Establish an optimization branch and freeze the search context and sorting position of the initial search document list; select a target document from the initial search document list and replace it with the optimized document variant to form a revised document list; input the revised document list into the target generative engine to generate an optimized response;

[0019] By comparing the differences between the optimized response and the baseline response, the causal impact of changes in the target document content on the generated results is quantified.

[0020] As a preferred option, the dual-layer memory mechanism includes:

[0021] Step-level memory, as immediate working memory, stores positive segments and forbidden paths in the current query optimization session; positive segments are text segments that the evaluated agent judges to have improved the score, and forbidden paths are operation records that cause the semantic fidelity to fall below a preset threshold.

[0022] Creator-level memory, serving as a cross-instance persistent knowledge base, stores general optimization strategy rules;

[0023] When generating instructions, the planning agent retrieves high-scoring strategies from the creator-level memory that match the current query scenario and engine, and combines them with positive fragments from the step-level memory to construct instructions.

[0024] Preferably, S3 includes:

[0025] S31. The planning agent reads the lessons learned from the previous round of optimization in the current step-level memory and the successful patterns in the creator-level memory, and generates the current round of optimization instructions by combining the preference profile.

[0026] S32. The editing agent generates multiple candidate variants in parallel according to the optimization instructions of this round;

[0027] S33. The evaluation agent uses a two-stage selection mechanism: the candidate variant with the largest gain is selected as the optimal variant; the next round is based on the optimal variant for optimization.

[0028] S34. If the gain of consecutive preset number of iterations is less than or equal to zero, or if the preset maximum number of iterations is reached, then stop the optimization.

[0029] As a preferred option, the two-stage selection mechanism includes:

[0030] The first stage of fidelity gating calculates the semantic fidelity of candidate variants relative to the target document; if the semantic fidelity is less than a preset threshold, the candidate variant is rejected and the erroneous path is recorded in the step-level memory to prevent semantic drift.

[0031] The second stage of gain prediction uses a large language model to predict the overall score gain brought by the candidate variants through gating; the variant with the largest overall score gain is selected as the optimal variant.

[0032] Preferably, the evaluation agent uses the DSV-CF dual-axis evaluation framework to calculate the comprehensive score of candidate variants. :

[0033]

[0034] in, The apparent semantic visibility score is used to quantify the exposure intensity of the source document in the generated response. The intrinsic semantic influence score is used to quantify the underlying contribution of the source document to the generated response logic. AA represents the attribution accuracy, used to verify whether the claims in the generated response are truly derived from the source document. The value range is [0,1]. As a balancing parameter, used to adjust the weight of visibility and influence, it is preferably set to 0.5; This is the attribution penalty coefficient, used to penalize the total score when the attribution accuracy is low, preventing the model from producing illusions.

[0035] As a preferred option, the apparent semantic visibility score The calculation formula is as follows:

[0036]

[0037] in, These include word-level visibility (WLV), positional decay authority (DPA), citation salience (CP), and subjective impression (SI). Let i be the weight of the i-th sub-index. The normalization function maps the index to the [0,1] interval.

[0038] Preferably, the overall score gain is the difference between the overall score obtained by the evaluation agent in the baseline branch and the overall score obtained by the evaluation agent in the optimization branch.

[0039] As a preferred option, in S4, the specific steps of performing closed-loop memory solidification include:

[0040] Determine whether the current optimization task generates positive gain; if it does, organize the step-level data, including successful modification operations and failed operations rejected by gating, in the current optimization task, and use a large language model to abstract the step-level data.

[0041] Secondly, the generative engine optimization system for memory enhancement and multi-agent collaboration includes:

[0042] The experimental environment construction module is used to build a dual-branch control experimental environment and freeze the search context to isolate sorting interference;

[0043] The multi-agent collaboration module includes a preference agent, a planning agent, an editing agent, and an evaluation agent; the planning agent is responsible for formulating strategies, the editing agent is responsible for implementing modifications, and the evaluation agent is responsible for quality gating and scoring.

[0044] A dual-layer memory management module is used to maintain step-level memory and creator-level memory, enabling instant reuse and cross-task migration of strategies;

[0045] The optimization execution module is used to drive multiple rounds of evolutionary optimization iterations. In each round, the above module is called to generate the optimal document variant until the termination condition is met.

[0046] A generative engine optimization system for memory enhancement and multi-agent collaboration is used to execute generative engine optimization methods and steps such as those described in the first aspect.

[0047] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0048] (1) Scientific quantification of causal effects: This invention creatively adopts the strategies of "in-situ replacement" and "frozen context" through a two-branch control experimental protocol, effectively eliminating noise interference caused by fluctuations in search engine ranking algorithms, and for the first time realizing accurate causal measurement of content optimization effects in a black-box environment.

[0049] (2) Balancing high exposure and high credibility: The DSV-CF dual-axis evaluation system proposed in this invention introduces attribution accuracy as the core penalty item, and is combined with the fidelity gating mechanism of the evaluation agent. This fundamentally solves the problem that existing methods sacrifice content authenticity and generate illusory references in pursuit of high exposure, ensuring that the optimized content has both traffic and credibility.

[0050] (3) Possesses lifelong learning capabilities: Through creator-level memory, this system can abstract successful optimization experiences for a single query into general natural language rules (such as format preferences for a specific engine). As the number of processing tasks increases, the system can automatically accumulate knowledge, avoiding the "catastrophic forgetting" of traditional methods and significantly improving the optimization efficiency of new tasks.

[0051] (4) Engine preference adaptation: The preference agent of this invention can construct a profile of a specific engine and combine it with the strategy generation of the planning agent, so that the optimization results can accurately adapt to the differentiated needs of different generative engines and have stronger generalization ability. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method framework of Embodiment 1 of the present invention. Detailed Implementation

[0053] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0054] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0055] Example 1: Optimizing healthcare queries for the Gemini engine

[0056] like Figure 1 The method shown is a generative engine optimization method combining memory enhancement and multi-agent collaboration. The method mainly includes the following steps:

[0057] S1. Obtain the user query and initial retrieval document list of the target generative engine; construct a two-branch control experimental environment including a baseline branch and an optimization branch;

[0058] In this embodiment, Gemini-3 Pro is selected as the target generative engine, and Tavily is selected as the simulated retrieval service provider.

[0059] The specific construction process is as follows:

[0060] (1) Input capture: The system receives user query q (e.g., "What are the side effects of taking aspirin?").

[0061] (2) Initial search: Call the Tavily Search API to obtain the list of Top-K documents that are most relevant to the user's query q. In this embodiment, K=10 is set as the initial list of documents to be searched.

[0062] (3) Establishment of two branches:

[0063] Baseline branch (Branch 1): Maintains the Top-K document list Without changing it, use it as context input to Gemini-3 Pro and record the generated benchmark response. .

[0064] Optimize branch (Branch 2): From the Top-K document list Uniformly sample one document as the target document This invention employs an "in-situ replacement" strategy, that is, the optimized document... Replace the target document in the list Get a list of corrected documents Note that, except Apart from content changes, the content of other documents in the list and the sort order of all documents remain strictly frozen. This operation ensures that any subsequent observed changes in the generated results are entirely attributable to the optimized documents. The quality of the content, rather than the random fluctuations of search engine ranking algorithms.

[0065] S2. Construct and initialize a virtual editing team that includes a preference agent, a planning agent, an editing agent, and an evaluation agent; the preference agent constructs a preference profile for a specific generative engine based on large-scale historical data;

[0066] like Figure 1 As shown, the agents in the virtual editing team interact with each other through the connections illustrated. This invention constructs four agents with distinct functions based on Gemini. To ensure the agents can accurately execute tasks while maintaining the confidentiality of the core design, this embodiment partially discloses the core prompt design for each agent:

[0067] (1) Preference Agent ):

[0068] This intelligent agent does not participate in real-time optimization loops, but instead analyzes large-scale offline "query-source-response" triple data to construct a preference profile for the target engine. ).

[0069] Prompt template:

[0070] "You are a data analyst. Please analyze the following search logs for [target engine name]. Compile and summarize the content characteristics that this engine tends to cite when generating responses. Output a profile from the following dimensions." :

[0071] Source preference (e.g., do you prefer .gov or .edu domains?);

[0072] Content DNA (e.g., average sentence length of frequently cited paragraphs, number density, punctuation habits);

[0073] Formatting characteristics (e.g., preference for Markdown tables or lists).

[0074] The output should be a structured natural language description.

[0075] (2) Planning Agent ):

[0076] As the "editor-in-chief," one is responsible for developing revision strategies rather than directly modifying the text. This incorporates preference profiling. Historical memory and current step memory Generate optimization instructions.

[0077] Prompt template:

[0078] "You are a professional GEO editor. Your goal is to improve the visibility of documents in the [Target Engine]'s generated results while maintaining the authenticity of the content."

[0079] Input information:

[0080] Engine profile: [Insert] ]

[0081] Long-term success strategy: [Insert from historical memory] [Relevant rules retrieved]

[0082] Feedback from the current attempt: [Insert current step memory] [Failure / Success Records]

[0083] Original document snippet: [Insert document]

[0084] Task: Please specify the modification instructions.

[0085] Require:

[0086] Instructions must be specific (e.g., 'Add statistics to the second paragraph', not 'Make the article better').

[0087] Strictly adhere to safety regulations and do not request the falsification of facts.

[0088] Optimize to address weaknesses identified in the engine profile.

[0089] (3) Editor Agent ):

[0090] As the "executor," it generates specific optimized document variants based on optimization instructions.

[0091] Prompt template:

[0092] "You are a skilled editor. Please revise the given document according to the following instructions."

[0093] Command: [Insert] [Generated instructions]

[0094] Original document: [Insert] ]

[0095] Output requirements:

[0096] Output K different modified versions (K=3 in this example).

[0097] The core semantics of the original document remain unchanged.

[0098] The modifications appear natural and conform to human language habits.

[0099] (4) Evaluator Agent ):

[0100] As a "quality inspector", I used the LLM-as-a-Judge strategy to score.

[0101] Prompt template:

[0102] "You are a fair reviewer. Please evaluate the optimized document." Compared with the original document The relationship between them is analyzed, and their performance in generative engines is predicted.

[0103] Step 1: Fidelity Check

[0104] Judge the optimized document Did it remain? The core facts. Output semantic fidelity. Semantic fidelity is the result of scoring a large model by defining cue words. If... (threshold) If it is, then it is marked as 'reject'.

[0105] Step 2: Influence Prediction

[0106] Predicted optimized documents Potential improvement in visibility. Please rate the following dimensions (0-10 points):

[0107] Citation salience (CP): Does the citation appear in a prominent position?

[0108] Attribution accuracy (AA): Does the content support the claim?

[0109] Output: A rating report in JSON format.

[0110] S3. Use a virtual editing team to perform multiple rounds of evolutionary optimization iterations in the optimization branch to generate the final optimized document. In each iteration, the planning agent combines preference profiles and a two-layer memory mechanism to generate optimization instructions, the editing agent generates multiple candidate variants according to the optimization instructions, and the evaluation agent screens and scores the candidate variants based on a two-axis semantic metric, and selects the best variant as the benchmark for the next round of optimization.

[0111] The dual-layer memory mechanism includes:

[0112] Step-level memory As immediate working memory, it stores positive segments and forbidden paths in the current query optimization session; positive segments are text segments that the evaluation agent determines have improved the score, and forbidden paths are those that cause document fidelity to fall below a preset threshold. Operation logs;

[0113] Creator-level memory ): Serves as a cross-instance persistent knowledge base, storing general optimization strategy rules;

[0114] When generating instructions, the planning agent retrieves high-scoring strategies from the creator-level memory that match the current query scenario and engine, and combines them with positive fragments from the step-level memory to construct instructions.

[0115] This step is a cyclical process, specifically including:

[0116] S31, Instruction Generation: Planning the Intelligent Agent Read current step level memory Lessons learned from previous failures (e.g., "Don't overuse adjectives, or it will be flagged as an ad"), combined with creator-level memory. The successful patterns in the process generate the optimization instructions for this round.

[0117] S32, Parallel Candidate Generation: Editing the Intelligent Agent Three candidate variants are generated in parallel according to the same instruction. .

[0118] S33, Two-phase selection mechanism:

[0119] The first stage is the Fidelity Gate: calculating the semantic fidelity of candidate variants relative to the target document. ;like If the candidate variant is rejected, it will be rejected directly to prevent semantic drift.

[0120] The second stage is gain prediction: using a large language model to predict the potential gains of candidate variants passed through the gating mechanism.

[0121] Gain: Select the variant with the largest predicted gain as the optimal variant.

[0122] In this embodiment:

[0123] Phase 1 (Fidelity Gating): Evaluating the agent's computation of each candidate variant. Assuming The semantics have undergone a significant shift (e.g., changing "mild side effects" to "no side effects"), resulting in If the candidate variant is not found, it is discarded and the erroneous path is recorded in the step-level memory. As a "forbidden zone".

[0124] Second stage (gain prediction): For the candidate variants that pass the gate, calculate the overall score gain of their predictions. The calculation method involves subtracting the scores obtained from the baseline branch's evaluation agent and the scores obtained from the optimization branch's evaluation agent. The variant with the largest gain is selected as the baseline for the next round. That is, the next stage optimizes on the variant that has the greatest gain in this stage.

[0125] The evaluation agent employs the DSV-CF (Dual-Axis Semantic Visibility & Content Fidelity) dual-axis evaluation framework to calculate the comprehensive score of candidate variants.

[0126]

[0127] in:

[0128] The surface semantic visibility score is used to quantify the exposure intensity of the source document in the generated response.

[0129] The Intrinsic Semantic Impact score quantifies the underlying contribution of the source document to the generated response logic.

[0130] AA stands for Attribution Accuracy, which is used to verify whether the claims in the generated response are truly derived from the source document. Its value ranges from [0,1].

[0131] As a balancing parameter, used to adjust the weight of visibility and influence, it is preferably set to 0.5;

[0132] This is the attribution penalty coefficient, used to penalize the total score when the attribution accuracy is low, preventing the model from producing illusions.

[0133] Apparent semantic visibility score The calculation formula involves weighted aggregation of the following sub-indicators:

[0134]

[0135] in, These include word-level visibility (WLV), positional decay authority (DPA), citation salience (CP), and subjective impression (SI). Let be the weight of the i-th sub-index; The normalization function maps the index to the [0,1] interval.

[0136] Overall Score The calculation is specified in this embodiment as follows:

[0137]

[0138] in, Set it to 0.5 to balance visibility and influence. The threshold is set to 2.0 to severely punish hallucinations. WLV (word-level visibility) and DPA (positional decay authority) are calculated through rule matching, while other metrics (CP, SI, AA, etc.) are obtained by evaluating agent scores.

[0139] S34. Termination condition: If the gain of consecutive iterations P = 3 rounds is... If the maximum number of iterations is reached (e.g., 10 rounds), then the optimization stops.

[0140] S4. After the optimization task is completed, execute closed-loop memory solidification; abstract the final optimization document optimization trajectory into natural language rules and store it in creator-level memory to update the dual-layer memory mechanism.

[0141] After the optimization task for query q is completed, if the final optimized document A positive gain was achieved compared to the initial document. The system performs the following memory update operations:

[0142] (1) Step-by-step experience archiving: Archive the experience gained in this session The data (including successful modification operations and failed operations rejected by the gating system) is organized.

[0143] (2) Creator-level pattern aggregation: Utilizing large language models to... Abstraction.

[0144] Memory reinforcement prompt template:

[0145] "Please analyze the operation log of this optimization task."

[0146] Scenario: [Query type: Medical], [Engine: Gemini]

[0147] Successful operation: [For example: converting a long paragraph into a Markdown table and adding a data source column to the table header]

[0148] Failed operation: [For example: adding a large number of repeated keywords at the end]

[0149] Task: Please abstract the above experience into a general optimization rule.

[0150] Output example: 'When the Gemini engine processes medical queries, using structured tables and explicitly annotating data sources can significantly improve citation probability; keyword stuffing should be avoided.'

[0151] The generated natural language rules are stored in the creator-level memory constructed by the vector database Milvus. In the next instance of a similar medical query, the planning agent will prioritize recalling the rule via vector retrieval.

[0152] Example 2:

[0153] A generative engine optimization system for memory enhancement and multi-agent collaboration includes:

[0154] The experimental environment construction module is used to build a dual-branch control experimental environment and freeze the search context to isolate sorting interference;

[0155] This module is deployed on a cloud server and connects to the target generative engine (GPT-5.1 API, Gemini-3 pro, Qwen-3 max) and the search engine (Tavily API) via API interfaces. The module embeds a "context freeze controller" to lock all search results except the target document in the optimization branch, ensuring the controllability of the experimental environment.

[0156] The multi-agent collaboration module includes a preference agent, a planning agent, an editing agent, and an evaluation agent; the planning agent is responsible for formulating strategies, the editing agent is responsible for implementing modifications, and the evaluation agent is responsible for quality gating and scoring.

[0157] This module contains four sub-modules, each corresponding to one of the four intelligent agents:

[0158] Preference analysis unit: runs periodically, processes log data offline, and updates. .

[0159] Strategy Planning Unit: Loads the planning agent model and is responsible for the dynamic assembly and reasoning of prompt words.

[0160] Content editing unit: Loads and edits intelligent agent models, supports concurrent calls to achieve parallel generation of candidate variants.

[0161] Evaluation Gating Unit: Loads the evaluation agent model and embeds logic judgment circuits for execution. A hard block.

[0162] A dual-layer memory management module is used to maintain step-level memory and creator-level memory, enabling real-time reuse and cross-task migration of strategies; including:

[0163] Short-term memory (RAM / Redis): used for storage It supports high-frequency read and write operations and records the current session state.

[0164] Long-term memory area (Vector Database): Used to store... It transforms natural language rules into high-dimensional vectors through the Embedding model, supporting fast retrieval based on semantic similarity (Top-K Retrieval).

[0165] The optimization execution module drives multiple rounds of evolutionary optimization iterations. In each round, it calls the modules mentioned above to generate the optimal document variant until the termination condition is met. As the system's central control unit, it is responsible for scheduling these modules. It includes an "evolutionary controller" for monitoring the number of iteration rounds and calculating... The score changes, the termination condition is determined, and the memory solidification process is triggered.

[0166] A generative engine optimization system for memory enhancement and multi-agent collaboration is used to execute generative engine optimization methods and steps such as those described in the first aspect.

[0167] The workflow of this system is as follows: user input query -> experimental environment construction module initializes dual branches -> optimization execution module starts iteration -> planning / editing / evaluation units work together -> memory management module provides strategy support and records new knowledge -> outputs the optimized document content and its corresponding optimization report.

[0168] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A generative engine optimization method based on memory enhancement and multi-agent collaboration, characterized in that, Includes the following steps: S1. Obtain the user query and initial retrieval document list of the target generative engine; construct a two-branch control experimental environment including a baseline branch and an optimization branch; S2. Construct and initialize a virtual editing team that includes a preference agent, a planning agent, an editing agent, and an evaluation agent; Preference agents construct preference profiles for specific generative engines based on large-scale historical data; S3. Use a virtual editing team to perform multiple rounds of evolutionary optimization iterations in the optimization branch to generate the final optimized document; In each iteration, the planning agent generates optimization instructions by combining preference profiles and a two-layer memory mechanism, the editing agent generates multiple candidate variants based on the optimization instructions, and the evaluation agent filters and scores the candidate variants based on a two-axis semantic metric, selecting the best variant as the benchmark for the next round of optimization. S4. After the optimization task is completed, execute closed-loop memory solidification; abstract the final optimization document optimization trajectory into natural language rules and store it in creator-level memory to update the dual-layer memory mechanism.

2. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 1, characterized in that, In S1, the construction of the two-branch control experimental environment includes: Establish a baseline branch, keeping the original order and content of the initial searched document list unchanged, and input the target generative engine to generate a baseline response; Establish an optimization branch and freeze the search context and sorting position of the initial search document list; select a target document from the initial search document list and replace it with the optimized document variant to form a revised document list; input the revised document list into the target generative engine to generate an optimized response; By comparing the differences between the optimized response and the baseline response, the causal impact of changes in the target document content on the generated results is quantified.

3. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 1, characterized in that, The dual-layer memory mechanism includes: Step-level memory, as immediate working memory, stores positive segments and forbidden paths in the current query optimization session; positive segments are text segments that the evaluated agent judges to have improved the score, and forbidden paths are operation records that cause the semantic fidelity to fall below a preset threshold. Creator-level memory, serving as a cross-instance persistent knowledge base, stores general optimization strategy rules; When generating instructions, the planning agent retrieves high-scoring strategies from the creator-level memory that match the current query scenario and engine, and combines them with positive fragments from the step-level memory to construct instructions.

4. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 1, characterized in that, S3 include: S31. The planning agent reads the lessons learned from the previous round of optimization in the current step-level memory and the successful patterns in the creator-level memory, and generates the current round of optimization instructions by combining the preference profile. S32. The editing agent generates multiple candidate variants in parallel according to the optimization instructions of this round; S33. The evaluation agent uses a two-stage selection mechanism: the candidate variant with the largest gain is selected as the optimal variant; the next round is based on the optimal variant for optimization. S34. If the gain of consecutive preset number of iterations is less than or equal to zero, or if the preset maximum number of iterations is reached, then stop the optimization.

5. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 4, characterized in that, The two-phase selection mechanism includes: The first stage of fidelity gating calculates the semantic fidelity of candidate variants relative to the target document; if the semantic fidelity is less than a preset threshold, the candidate variant is rejected and the erroneous path is recorded in the step-level memory to prevent semantic drift. The second stage of gain prediction uses a large language model to predict the overall score gain brought by the candidate variants through gating; the variant with the largest overall score gain is selected as the optimal variant.

6. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 5, characterized in that, The agent is evaluated using the DSV-CF dual-axis evaluation framework to compute the comprehensive score of candidate variants. : ； in, The apparent semantic visibility score is used to quantify the exposure intensity of the source document in the generated response. The intrinsic semantic influence score is used to quantify the underlying contribution of the source document to the generated response logic. AA represents the attribution accuracy, used to verify whether the claims in the generated response are truly derived from the source document. The value range is [0,1]. As a balancing parameter, used to adjust the weight of visibility and influence, it is preferably set to 0.5; This is the attribution penalty coefficient, used to penalize the total score when the attribution accuracy is low, preventing the model from producing illusions.

7. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 6, characterized in that, Apparent semantic visibility score The calculation formula is as follows: ； in, These include word-level visibility (WLV), positional decay authority (DPA), citation salience (CP), and subjective impression (SI). Let i be the weight of the i-th sub-index. The normalization function maps the index to the [0,1] interval.

8. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 5, characterized in that, The overall score gain is the difference between the overall score obtained by the evaluation agent in the baseline branch and the overall score obtained by the evaluation agent in the optimization branch.

9. The generative engine optimization method for memory enhancement and multi-agent collaboration according to claim 1, characterized in that, In S4, the specific steps for executing closed-loop memory solidification include: Determine whether the current optimization task generates positive gain; if it does, organize the step-level data, including successful modification operations and failed operations rejected by gating, in the current optimization task, and use a large language model to abstract the step-level data.

10. A generative engine optimization system for memory enhancement and multi-agent collaboration, characterized in that, include: The experimental environment construction module is used to build a dual-branch control experimental environment and freeze the search context to isolate sorting interference; The multi-agent collaboration module includes a preference agent, a planning agent, an editing agent, and an evaluation agent; the planning agent is responsible for formulating strategies, the editing agent is responsible for implementing modifications, and the evaluation agent is responsible for quality gating and scoring. A dual-layer memory management module is used to maintain step-level memory and creator-level memory, enabling instant reuse and cross-task migration of strategies; The optimization execution module is used to drive multiple rounds of evolutionary optimization iterations. In each round, the above module is called to generate the optimal document variant until the termination condition is met. The memory-enhanced and multi-agent collaborative generative engine optimization system is used to execute the memory-enhanced and multi-agent collaborative generative engine optimization method and steps as described in claim 1.

Images