Optimization methods, devices, and media based on long-term capability evolution trajectories

By constructing a memory knowledge graph and tracing back to the source, the shortcomings of evaluating the long-term capability evolution trajectory of intelligent agents are solved, achieving a comprehensive, realistic, and interpretable reflection and optimization of the capabilities of intelligent agents, and improving iteration efficiency.

CN123070263APending Publication Date: 2026-11-10FANXING INTELLIGENT COMPUTING TECHNOLOGY (BEIJING) CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611160595.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-31
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies are unable to fully, realistically, and interpretably reflect the long-term capability evolution trajectory of intelligent agents. They lack dynamic tracking of the entire chain, including historical behavioral patterns, experience accumulation, error correction, and knowledge updates. This results in evaluation results lacking continuity and correlation over time, making it difficult to support iterative optimization of intelligent agents.

Method used

By constructing a memory knowledge graph, we can determine the quantitative values ​​of the evolutionary dimensions and the comprehensive evolutionary score of the memory agent, generate evolutionary trajectories and trace the relationships, optimize the memory agent, and realize the structured processing and source analysis of memory knowledge data.

Benefits of technology

It achieves a comprehensive, realistic, and interpretable reflection of the long-term capability evolution process of intelligent agents, improves the efficiency and interpretability of iterative optimization, and enhances the traceability and verifiability of evaluation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070263A_ABST
    Figure CN123070263A_ABST
Patent Text Reader

Abstract

This application provides an optimization method, apparatus, and medium based on long-term capability evolution trajectory. By determining the memory knowledge data of the memory agent and performing structured processing, a corresponding memory knowledge graph is obtained to determine the quantified value of the evolutionary dimension and the comprehensive evolutionary score corresponding to the memory agent. According to a preset time granularity, the quantified value of the evolutionary dimension, the comprehensive evolutionary score, and the update information of the memory knowledge graph corresponding to each time granularity node are determined as feature labels for each time granularity node. Based on the time sequence, the evolutionary trajectory of each feature label changing over time is obtained, and a source-tracing association relationship between the evolutionary trajectory and the memory knowledge graph is established. Mutation nodes and corresponding mutation events are determined based on the source-tracing association relationship, and the memory agent is optimized based on the mutation events. According to the method provided in this application, the memory agent can be optimized through the memory knowledge graph and the evolutionary trajectory.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and in particular relates to an optimization method, apparatus and medium based on long-term capability evolution trajectory. Background Technology

[0002] Intelligent agents can monitor the full lifecycle capabilities and autonomously evolve and optimize large language models and multi-agent collaborative systems, thereby achieving deep coupling analysis of the evolution of agent capabilities and the accumulation of memorized knowledge.

[0003] In existing technologies, static snapshot-style evaluation of the performance of a single task can be conducted, with numerical statistics performed on a single dimension such as task completion accuracy, response latency, and tool call efficiency. Advanced solutions can combine the completion status of multiple rounds of tasks within a limited time window to perform short-term capability rating. In terms of memory management, linear serialization storage can be used to achieve raw storage and keyword matching retrieval.

[0004] However, existing capabilities for tracking the long-term evolution of intelligent agents lack dynamic tracking of the entire chain, including historical behavioral patterns, experience accumulation, error correction, and knowledge updates, making it difficult to construct a complete long-term evolution path. Summary of the Invention

[0005] This application provides an optimization method, apparatus, and medium based on long-term capability evolution trajectory, which can generate a comprehensive, realistic, and interpretable reflection of the dynamic evolution process of the capabilities of a memory agent through a memory knowledge graph.

[0006] In a first aspect, embodiments of this application provide an optimization method based on a long-term capability evolution trajectory, the method comprising: The memory knowledge data of the memory agent is determined, and the memory knowledge data is structured according to the preset graph construction rules to obtain the memory knowledge graph corresponding to the memory agent. Based on the memory knowledge graph, the evolution dimension quantification value corresponding to the memory agent is determined, and the comprehensive evolution score is determined based on the evolution dimension quantification value; According to the preset time granularity, the evolution dimension quantification value, comprehensive evolution score and memory knowledge graph update information corresponding to each time granularity node are determined as the feature labels of each time granularity node; Based on the temporal sequence of time granularity nodes, the evolution trajectory of each feature label over time is obtained, and the tracing relationship between the evolution trajectory and the memory knowledge graph is established. If the change in the comprehensive evolution score between adjacent time granularity nodes exceeds a preset change threshold, the corresponding time granularity node will be identified as a mutation node. Based on the source-tracing relationship, the mutation events corresponding to the mutation nodes are identified on the memory knowledge graph, and the memory agent is optimized based on the mutation events.

[0007] Secondly, embodiments of this application provide an optimization apparatus based on a long-term capability evolution trajectory, the apparatus comprising: The module is used to determine the memory knowledge data of the memory agent, and to perform structured processing on the memory knowledge data according to the preset graph construction rules to obtain the memory knowledge graph corresponding to the memory agent. The processing module is used to determine the evolution dimension quantification value corresponding to the memory agent based on the memory knowledge graph, and to determine the comprehensive evolution score based on the evolution dimension quantification value; The feature module is used to determine the evolution dimension quantification value, comprehensive evolution score and memory knowledge graph update information corresponding to each time granularity node as the feature label of each time granularity node according to the preset time granularity. The association module is used to obtain the evolution trajectory of each feature label over time based on the temporal sequence of time granularity nodes, and to establish the source association relationship between the evolution trajectory and the memory knowledge graph; The mutation module is used to identify the corresponding time granularity node as a mutation node when the change in the comprehensive evolution score between adjacent time granularity nodes exceeds a preset change threshold. The optimization module is used to determine the mutation events corresponding to mutation nodes on the memory knowledge graph based on the source-tracing relationship, and optimize the memory agent based on the mutation events.

[0008] Thirdly, embodiments of this application provide a terminal device, the device including: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the optimization method based on long-term capability evolution trajectory as described in the first aspect.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the optimization method based on long-term capability evolution trajectory as described in the first aspect.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed, implements the optimization method based on long-term capability evolution trajectory as described in the first aspect.

[0011] The optimization method, apparatus, and medium based on long-term capability evolution trajectory of this application embodiment determine the memory knowledge data of the memory agent, perform structured processing on the memory knowledge data according to preset graph construction rules, and obtain the memory knowledge graph corresponding to the memory agent. This enables the structured representation and storage of the memory knowledge data in the memory agent, facilitating the sorting of relevant data through the memory knowledge graph, and thus obtaining the long-term capability evolution trajectory. Based on the memory knowledge graph, the evolution dimension quantification value corresponding to the memory agent is determined, and a comprehensive evolution score is determined based on the evolution dimension quantification value. This allows for the determination of the long-term evolution capability of the memory agent from multiple dimensions. According to a preset time granularity, the evolution dimension quantification value, comprehensive evolution score, and update information of the memory knowledge graph corresponding to each time granularity node are determined as the features of each time granularity node. Tags can display the evolution of the memory agent's capabilities over time. Based on the temporal sequence of time-granularity nodes, the evolutionary trajectory of each feature tag over time is obtained, and a source-tracing relationship is established between the evolutionary trajectory and the memory knowledge graph. After sorting each feature tag according to the temporal sequence, the evolutionary trajectory is obtained, and a source-tracing relationship is established. This can associate points in the evolutionary trajectory with events on the memory knowledge graph, facilitating source tracing and improving interpretability. Based on the source-tracing relationship, the mutation events corresponding to mutation nodes are determined on the memory knowledge graph. The memory agent is optimized based on the mutation events. Thus, the memory knowledge graph can couple the memory agent's memory knowledge accumulation, experience reuse, knowledge system update, and error iteration update, providing a more comprehensive, realistic, and interpretable reflection of the long-term capability evolution trajectory of the memory agent and enabling its optimization. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an optimization method based on a long-term capability evolution trajectory provided in an embodiment of this application; Figure 2 This is a schematic diagram of the architecture of a long-term capability evolution trajectory assessment system provided in an embodiment of this application; Figure 3 This is a schematic diagram of an optimization device structure based on a long-term capability evolution trajectory provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0014] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0016] It should be noted that the acquisition, storage, use, and processing of data in this application embodiment all comply with the relevant provisions of laws and regulations.

[0017] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0018] In existing technologies, the evaluation techniques for intelligent agents' capabilities mostly focus on static snapshot-style evaluation of the performance of a single task. They mainly conduct numerical statistics around a single dimension such as task completion accuracy, instruction response latency, and tool call efficiency. Some advanced solutions combine the completion status of multiple rounds of tasks within a limited time window to carry out short-term capability rating.

[0019] In terms of agent memory management, existing technologies mainly use linear serialization storage to record historical decision-making behavior, human-computer interaction data, and knowledge update content of agents. This achieves the raw storage of memory data and simple keyword matching retrieval. However, it does not perform in-depth processing such as deep structured analysis, multi-dimensional association modeling, and semantic layer fusion on the memory data. For the long-term capability evolution of agents, existing technologies lack full-link dynamic tracking of their historical behavior patterns, experience accumulation effects, error iteration correction, and knowledge system updates. Simple numerical comparisons of static evaluation results at different time points make it difficult to construct a complete evolution path of agent capabilities from basic execution to high-level applications.

[0020] Meanwhile, the existing evaluation system has not accurately quantified and modeled dimensions that are strongly correlated with capability evolution, such as the adaptability of the agent's memory knowledge retrieval, the proportion of redundancy in the knowledge system, and the effective memory decay rate. It also cannot extract the intrinsic driving laws of the agent's capability improvement and the explicit shortcomings of capability development through in-depth mining of historical memory knowledge data. Furthermore, it has not established a quantitative correlation model between memory knowledge accumulation and capability evolution, making it difficult for the evaluation results to support the positive optimization of the agent.

[0021] Furthermore, some existing evaluation schemes attempt to bypass the deep processing of memory data and directly evaluate the capabilities of intelligent agents, such as by inferring capability status from task results. However, this lacks support for the underlying data of capability evolution, resulting in evaluation results that lack traceability and interpretability, making it difficult to pinpoint the core reasons for capability changes and providing targeted basis for the iterative optimization of intelligent agents.

[0022] In other words, in existing technologies, the evaluation dimension of the long-term capability evolution trajectory of intelligent agents has a single static feature, focusing on the surface results of task execution. It is easy to ignore the core elements of capability evolution such as experience accumulation, error iteration and correction, and knowledge system evolution in the long-term operation of intelligent agents. As a result, it is difficult to reflect the dynamic evolution process of the capability of intelligent agents from quantitative to qualitative change, and the evaluation results lack the continuity and correlation of the time dimension.

[0023] Furthermore, the linear serialization storage mode of intelligent agent memory data lacks multi-dimensional structured association and semantic layer fusion modeling, making it difficult to explore the inherent logical connections and causal relationships between historical memory data. As a result, memory data is difficult to effectively support the source analysis and root cause localization of ability evolution. At the same time, the evaluation results of long-term ability evolution trajectory lack scientificity and quantification, which leads to low iteration efficiency of intelligent agents and blind optimization direction.

[0024] To address the problems in the prior art, embodiments of this application provide an optimization method, apparatus, and medium based on long-term capability evolution trajectories.

[0025] The optimization method based on long-term capability evolution trajectory provided in the embodiments of this application will be introduced first.

[0026] Figure 1 A flowchart illustrating an embodiment of the optimization method based on long-term capability evolution trajectory provided in this application is shown. Figure 1 As shown, the method may include the following steps: S101: Determine the memory knowledge data of the memory agent, perform structured processing on the memory knowledge data according to the preset graph construction rules, and obtain the memory knowledge graph corresponding to the memory agent.

[0027] Specifically, the memory knowledge data generated by the memory agent during its operation is determined. This memory knowledge data can include interaction records, task results, decision logs, error samples, and reuse experience. For example, taking a customer service agent as an example, its memory knowledge data can include historical dialogue records with users, problem-solving paths, successful and unsuccessful handling plans, and response strategies in different business scenarios. The memory knowledge data can be the raw data of the entire operation cycle of the memory agent.

[0028] In addition, according to the preset graph construction rules, the memory knowledge data is processed in a structured manner to obtain the memory knowledge graph corresponding to the memory agent. The memory knowledge graph is a structured long-term memory storage carrier for the memory agent. It can organize the memory knowledge data in a graph structure of entities, relationships and attributes, and can be used for association retrieval and topology analysis. The preset graph construction rules can be an evolvable weighted knowledge graph model.

[0029] This application embodiment constructs a memory knowledge graph from memory knowledge data to transform the memory knowledge data into a weighted structured graph, and presents the relationship between memory knowledge data in the memory agent through the memory knowledge graph.

[0030] S102: Based on the memory knowledge graph, determine the evolution dimension quantification value corresponding to the memory agent, and determine the comprehensive evolution score based on the evolution dimension quantification value.

[0031] Specifically, based on the memory knowledge graph, the evolution dimension quantification value corresponding to the memory agent is determined. The evolution dimension quantification value can include the ability evolution rate, memory knowledge retrieval adaptability, knowledge system redundancy ratio, effective memory decay rate, and scenario generalization and expansion capability.

[0032] Among them, the capability evolution rate is used to calculate the improvement in the completion rate of similar tasks within a fixed window; the memory knowledge retrieval adaptation rate is used to calculate the success rate of the retrieval of graph nodes in new tasks; the knowledge system redundancy ratio is used to calculate the ratio of the weight of nodes corresponding to the three types of events (unreused, duplicate, and conflict) to the total weight; and the effective memory decay rate is used to calculate the proportion of the weighted decay of unreused effective nodes to the total weight of effective nodes within the life cycle.

[0033] In addition, the comprehensive evolution score can be determined by weighted summation of the quantitative values ​​of each evolution dimension, with the sum of the weight coefficients of each evolution dimension being 1.

[0034] Among these features, a dynamic weight adjustment model can be constructed based on the application scenarios and core business needs of the memory agent, assigning personalized weight coefficients to the quantized values ​​of each evolution dimension. (satisfy The overall evolution score of the agent is obtained by weighted summation. The formula for calculating the overall evolution score can be expressed as: ; in, The overall evolution score ranges from 0 to 100. This represents the quantized value of the evolutionary dimension corresponding to the rate of capability evolution. This represents the evolutionary dimension quantification value corresponding to the adaptability of memory knowledge retrieval. This represents the quantitative value of the evolution dimension corresponding to the redundancy ratio of the knowledge system. This represents the quantized value of the evolution dimension corresponding to the effective memory decay rate. This represents the quantified value of the evolution dimension corresponding to the scenario generalization and expansion capability.

[0035] The embodiments of this application can determine the evolution dimension quantification value and the corresponding comprehensive evolution score of the memory agent from multiple dimensions, thereby intuitively presenting the long-term evolution trend of the memory agent and enhancing the versatility and flexibility of the long-term evolution trajectory.

[0036] S103: According to the preset time granularity, the evolution dimension quantification value, comprehensive evolution score and memory knowledge graph update information corresponding to each time granularity node are determined as the feature labels of each time granularity node.

[0037] Specifically, the evolution dimension quantification value and comprehensive evolution score are discretized and sampled according to the preset time granularity. The preset time granularity is a configurable parameter, such as hour, week, day, month, etc. Different time granularities correspond to different analysis scenarios.

[0038] Extract the evolution dimension quantification value, comprehensive evolution score, and update information of the memory knowledge graph corresponding to each time granularity node within the time period between the current time granularity node and the previous time granularity node, and determine the above content as the feature label of each time granularity node.

[0039] The updated information includes, but is not limited to, the number of newly added knowledge nodes, the number of deleted knowledge nodes or the identifiers of merged redundant knowledge nodes, and the weight changes of directed edges in the memory knowledge graph. In the time-granularity nodes, each feature label exists independently. That is, in a time-granularity node, there are simultaneously evolution dimension quantification values, comprehensive evolution scores and updated information.

[0040] The embodiments of this application can provide structured data anchors for the generation of evolutionary trajectories and the tracing of mutations by using the feature labels of time-granularity nodes, carrying capacity scores, and map status.

[0041] S104: Based on the temporal sequence of time-granularity nodes, obtain the evolution trajectory of each feature label over time, and establish the source association relationship between the evolution trajectory and the memory knowledge graph.

[0042] Specifically, based on the temporal sequence of time granularity nodes, with time as the horizontal axis and the value of each feature label at that time granularity node as the vertical axis, the evolution trajectory of each feature label is generated. The evolution trajectory can be displayed in layers to show the overall trend and individual changes.

[0043] Furthermore, establishing a tracing relationship between evolutionary trajectories and memory knowledge graphs allows for the generation and storage of index identifiers for snapshots of the corresponding memory knowledge graphs at each time-granularity node in the evolutionary trajectory. This enables the backtracking from any point on the evolutionary trajectory to the topological structure and graph data of the memory knowledge graph at the time that point was generated.

[0044] In the memory knowledge graph, each node update, directed edge weight update, and topology adjustment event can be recorded with a corresponding timestamp to reverse locate the position and score change of the event on the evolutionary trajectory, thus forming a many-to-many mapping from the evolutionary trajectory to the memory knowledge graph.

[0045] This application embodiment generates an evolutionary trajectory based on feature tags and determines the source association between the evolutionary trajectory and the memory knowledge graph. It can directly correlate the score of the evolutionary trajectory with the specific data changes in the memory knowledge graph. At the same time, it can reflect the events in the memory knowledge graph to the trend of the evolutionary trajectory, thereby entering the knowledge level for interpretation and improving the traceability, verifiability and interpretability of the ability assessment.

[0046] S105: If the change in the comprehensive evolution score between adjacent time granularity nodes is greater than the preset change threshold, the corresponding time granularity node will be identified as a mutation node.

[0047] Specifically, mutation detection is performed on the evolutionary trajectory, which means determining the amount of change in the comprehensive evolution score between adjacent time granularity nodes. If the amount of change is greater than a preset change threshold, the corresponding time granularity node is identified as a mutation node.

[0048] The preset change threshold can be adaptively adjusted based on the historical score fluctuations of the memory agent, or it can be configured according to needs.

[0049] Furthermore, the direction of mutation can be determined. A positive mutation direction indicates an improvement in the memory agent's capabilities, while a negative mutation direction indicates a decrease in the memory agent's capabilities.

[0050] The embodiments of this application determine mutation nodes by the change in the comprehensive evolution score between adjacent time granularity nodes and a preset change threshold, which can reduce subjectivity and improve interpretability and reproducibility.

[0051] S106: Based on the source-tracing association, determine the mutation events corresponding to mutation nodes on the memory knowledge graph, and optimize the memory agent based on the mutation events.

[0052] Specifically, based on the source-tracing relationship, the time interval corresponding to the mutation node is determined, all change events that occur within the time interval are identified on the memory knowledge graph, and mutation events are identified among the change events. Mutation events include, but are not limited to, batch import of knowledge nodes, occurrence of high-frequency error nodes, sudden increase in knowledge redundancy, and breakthroughs in cross-scenario reuse of experience nodes.

[0053] Furthermore, the memory agent can be optimized through mutation events. The contribution direction and contribution value of each mutation event can be determined, and the memory agent can be optimized based on the contribution direction and contribution value to obtain the optimized memory agent.

[0054] The embodiments of this application can identify mutation events by tracing the source of the correlation, thereby optimizing the memory agent based on the mutation events and improving the long-term growth efficiency of the memory agent.

[0055] This application embodiment constructs a memory knowledge graph closely linked to the memory agent. In determining the long-term evolutionary trajectory of the memory agent's capabilities, the memory knowledge graph serves as the data foundation and analysis carrier for coordinated linkage. All evaluation stages revolve around the memory knowledge graph, reducing the disconnect between evaluation and memory knowledge data. Simultaneously, the memory knowledge graph enables structured processing of memory knowledge data within the memory agent and facilitates source analysis of the evolutionary trajectory. This allows for data-driven continuous iteration of the memory agent, enhancing its capability evolution and providing a more comprehensive, realistic, and interpretable reflection of the dynamic evolution process of the memory agent's capabilities.

[0056] In some embodiments, the memory knowledge data is structured according to preset graph construction rules to obtain a memory knowledge graph corresponding to the memory agent, including: Acquire historical memory data of the memory agent at various historical time points. The historical memory data includes historical log data, cross-scenario task association records, decision path distribution data, and task execution result data. Based on historical memory data, state difference data for each historical time node is determined. The state difference data is used to represent the changes in knowledge state, experience structure, and decision path distribution of the memory agent at adjacent time nodes. Based on the state difference data corresponding to each historical time point, multiple knowledge nodes of the memory knowledge graph are generated. Based on historical memory data, the calling relationship, reuse relationship and causal relationship between each knowledge node are determined, and multiple directed edges connecting each knowledge node are generated to obtain the memory knowledge graph corresponding to the memory agent.

[0057] Specifically, the historical memory data generated or updated by the memory agent at each historical time node is obtained. Here, the historical time node represents the past time granularity node relative to the current evaluation time, and the historical memory data can be obtained by polling.

[0058] Historical memory data includes, but is not limited to, historical log data, cross-scenario task association records, decision path distribution data, and task execution result data.

[0059] Historical log data includes, but is not limited to, input and output text of a single interaction, memory retrieval records, context fragments, and entity extraction results, used to represent the basic operational information of the memory agent at each historical time point. Cross-scenario task association records include, but are not limited to, knowledge repeatedly referenced in different business scenarios, corresponding task identifiers, and scenario adaptation parameters, used to represent the migration and adaptation behavior of the memory agent between different business scenarios or different types of tasks. Decision path distribution data includes, but is not limited to, reasoning steps, branch selection, logical jump paths, and strategy preference distribution of a single task, used to describe the possible paths traversed by the memory agent and their probability distribution when faced with specific data. Task execution result data includes, but is not limited to, task completion degree, error type labels, user feedback scores, and delivery outputs, used to represent feedback information after the task is completed.

[0060] In addition, state difference data for each historical time point is determined based on historical memory data. State difference data represents the amount of state change that occurs between two adjacent historical time points. That is, state difference data includes the amount of knowledge state change, the amount of experience structure change, and the amount of decision path distribution change, thereby reducing the large amount of duplicate and redundant data generated by full storage and improving resource utilization.

[0061] In addition, based on the state difference data corresponding to historical time nodes, multiple knowledge nodes of the memory knowledge graph are generated. Among them, the knowledge node is the smallest entity unit in the memory knowledge graph that carries a single independent piece of knowledge or experience. The built-in attributes of the knowledge node include, but are not limited to, the dimension to which it belongs, timestamp, change type identifier and original semantic content. All knowledge nodes are written into the memory knowledge graph in chronological order to form a node set.

[0062] In addition, the relationships between knowledge nodes are determined, and directed edges are generated for knowledge nodes with relationships, so as to connect the knowledge nodes through directed edges and obtain the memory knowledge graph corresponding to the memory agent.

[0063] Among them, the relationships include calling relationships, reuse relationships and causal relationships. Calling relationships represent the sequential calling relationship of knowledge nodes in a single task execution. Reuse relationships correspond to the repeated reference of the same knowledge node in different scenarios or tasks. Directed edges can connect source scenario nodes and target scenario nodes. Causal relationships represent the supporting role of prior knowledge in decision-making paths for subsequent conclusions. The direction of the directed edge is the causal reasoning direction.

[0064] Furthermore, each directed edge is accompanied by an initial association strength value, which can be determined based on the frequency of association occurrence and the positive correlation between task completion and the degree of association strength.

[0065] In some embodiments, the state information of the memory agent can also be obtained through an event-triggered acquisition mechanism, that is, when the memory agent meets the triggering conditions, the memory knowledge data of the memory agent can be obtained.

[0066] Triggering conditions include: the change in the decision path distribution of the memory agent compared to the previous historical time point exceeds a preset decision change threshold; a new knowledge node is introduced into the memory agent's knowledge base; and the error type appears for the first time during the operation of the memory agent. Triggering conditions can be represented by a trigger function, which is: [ ; in, This indicates an indicator function; a value of 1 triggers data collection, otherwise a value of 0. This represents the decision path distribution vector at historical time point t. This represents the decision path distribution vector at time t-1. Indicates the threshold for decision change. This indicates the introduction of a new knowledge node. This indicates an error type mutation marker.

[0067] Furthermore, during the process of confirming the state difference data, the state of the memory agent at time t corresponding to the historical time node can be defined as: ; in, This represents the knowledge state vector at time t, obtained from statistics on entities, attributes, and confidence levels in historical log data. The experience structure at time t is obtained by aggregating the reuse relationships and adaptation rules in the cross-scenario task association records. This represents the decision path distribution at time t, which is obtained by statistically analyzing the branch ratio and selection frequency in the decision path distribution data.

[0068] Based on the state of the memory agent at time t and the states of the memory agent at adjacent times, the state difference data is obtained and can be represented as: ; in, Represents state difference data. This represents the state of the memory agent at time t. This represents the state of the memory agent at time t, the state of the previous time step.

[0069] The methods of constructing and updating memory knowledge graphs using event triggering conditions and state difference data are complementary. State difference data can be used to determine the quantitative changes in memory knowledge in the memory knowledge graph, while triggering conditions are used to construct and update the memory knowledge graph from the perspective of qualitative change. They can be used together to obtain quantitative data on the magnitude of changes in each dimension of the memory knowledge graph, as well as label information on the specific event types that trigger the changes.

[0070] Furthermore, historical memory data obtained through multi-granularity semantic encoding can be processed so that each knowledge node in the memory knowledge graph carries semantic labels at three granularities: token-level, task-level, and policy-level. Its ternary representation can be expressed as: ; in, This represents token-level semantic encoding used for fine-grained entity matching. This represents task-level semantic encoding used for contextualized retrieval. It represents the semantic encoding at the policy level, which is used for high-level policy reuse and decision derivation, so that a single knowledge unit can carry multi-dimensional semantic information at the same time, which is convenient for computational needs of different granularities.

[0071] This application embodiment constructs and updates a memory knowledge graph using state difference data, controls the growth of memory knowledge in the memory knowledge graph, and describes the behavioral profile of knowledge nodes from multiple dimensions by using relations, reuse relations, and causal relationships as directed edges of the memory knowledge graph, making the memory knowledge graph a dynamically structured network that can record memory knowledge in memory agents and improve the accuracy of evolutionary trajectories.

[0072] In some embodiments, based on historical memory data, the calling relationships, reuse relationships, and causal relationships between knowledge nodes are determined, and multiple directed edges connecting each knowledge node are generated to obtain a memory knowledge graph corresponding to the memory agent, including: The importance of each knowledge node is determined based on the total number of times each knowledge node is invoked in the corresponding historical log data, the total number of times it is reused across scenarios in the cross-scenario task association records, and the centrality weight value in the causal path corresponding to each causal relationship. The initial weight of each directed edge is determined based on the adjacent call time interval and call frequency between each knowledge node; Based on the interval between the last call time and the first preset time between each knowledge node, the initial weight is subjected to time decay processing to obtain decay weight; Based on the reuse frequency between knowledge nodes and the node importance of the corresponding knowledge nodes, the decay weights are reused and strengthened to obtain a dynamically evolving memory knowledge graph.

[0073] Specifically, the importance of each knowledge node is determined based on the total number of times it is invoked in the corresponding historical log data, the total number of times it is reused across scenarios in cross-scenario task association records, and the centrality weight value in the causal path corresponding to each causal relationship. The node importance can be expressed as: ; in, Indicates the importance of nodes. Indicates the total number of times it was called. This indicates the total number of times the device is reused across different scenarios. Represents the centrality weight value. , and The weight coefficients corresponding to each indicator are summed to 1. The node importance is used to characterize the actual value of a single knowledge node. The higher the value, the stronger the support of the knowledge node for the task execution and decision-making of the memory agent.

[0074] In addition, for any pair of knowledge nodes with a calling relationship, the time records of the calling of the pair of knowledge nodes in all historical tasks are extracted from the historical log data. The time interval between two adjacent calls and the number of calls per unit time are calculated. The directed edges between the knowledge nodes are assigned initial weights based on the call time interval and the call frequency.

[0075] Extract the last call time corresponding to each directed edge, calculate the interval between it and the first preset time, and perform time decay processing on the initial weight to obtain the decayed weight. The first preset time can be the current evaluation time, that is, the interval between the last call time and the current time. The initial weight is subjected to time decay processing according to the interval to obtain the decayed weight.

[0076] In addition, the decay weight is reinforced by reuse. The reuse reinforcement is based on the reuse status of each knowledge node in the cross-scenario task association record. Reuse means that the same knowledge node is repeatedly called in different task scenarios or different types of tasks. If the number of times it is reused in cross-scenario or cross-task situations is high, it means that the knowledge node has high general value and its knowledge has a high degree of abstraction and transferability.

[0077] The attenuation weight can be enhanced by reusing it according to the frequency of reuse. The frequency of reuse and the degree of enhancement are positively correlated, resulting in the attenuation weight after reuse enhancement.

[0078] In addition, edges with reduced weights after reuse and reinforcement can be filtered out when both node importance and edge weight are below a preset threshold, resulting in a dynamically evolving memory knowledge graph.

[0079] The process of determining the weights of directed edges in a knowledge graph can be represented as follows: ; in, The weights represent the weights of the knowledge graph. Indicates the initial weights. This represents the time decay coefficient used for time decay processing. Indicates time interval, This represents the strengthening coefficient used in the reuse strengthening process. This represents the frequency of calls to associated edges per unit of time. i represents the index number of the starting source node of the directed edge, i.e., the knowledge node at the starting end of a directed edge in the memory knowledge graph. j represents the index number of the ending target node of the directed edge, i.e., the knowledge node at the pointing end of the same directed edge in the memory knowledge graph.

[0080] This application's embodiments, through time decay processing, reuse enhancement processing, and node importance assessment, enable the memory knowledge graph to respond to changes in the timeliness of knowledge, while also possessing cross-scenario and cross-task transfer capabilities, thereby achieving the autonomous evolution of the memory knowledge graph.

[0081] In some embodiments, the evolutionary dimension quantization value corresponding to the memory agent is determined based on the memory knowledge graph, including: Based on the memory knowledge graph, node state data and node association data of the knowledge nodes corresponding to the memory agent are extracted. The node state data is used to represent the calling characteristics, reuse characteristics, temporal decay characteristics and content validity characteristics of each knowledge node, and the node association data is used to represent the association between each knowledge node and the task execution result and the association between each knowledge node and the usage scenario. Based on node status data and node association data, determine the quantitative value of each evolution dimension. The evolution dimension includes at least one of the following: capability evolution rate, memory knowledge retrieval adaptability, knowledge system redundancy ratio, effective memory decay rate, and scenario generalization and expansion capability.

[0082] Specifically, the memory knowledge graph is the only data foundation and computational carrier for long-term capability evolution. All dimensions of quantitative computation require calling the knowledge nodes, directed edges, and associated application data of the memory knowledge graph, thereby reducing the risk of bypassing the memory knowledge of the memory agent for long-term capability evolution assessment from a computational logic perspective.

[0083] Based on the memory knowledge graph, the node state data and node association data of the knowledge nodes corresponding to the memory agent are extracted. Among them, the node state data represents the calling characteristics, reuse characteristics, temporal decay characteristics and content validity characteristics of each knowledge node.

[0084] Among them, the call feature is used to represent the cumulative number of calls, call time interval distribution and peak call period of the knowledge node; the reuse feature is used to represent the number of times the knowledge node is referenced across different business scenarios and the number of covered scenarios; the time decay feature is used to store the last call time of the directed edge associated with the knowledge node and the current weight value after time decay; and the content validity feature is used to mark the effective status, conflict mark and confidence level of the knowledge node.

[0085] Node association data is used to represent the mapping relationship between each knowledge node and external dimensions, including the association between knowledge nodes and task execution results, and the association between knowledge nodes and usage scenarios.

[0086] In addition, based on node status data and node association data, the quantitative values ​​of each evolution dimension are determined. The evolution dimension includes at least one of the following: capability evolution rate, memory knowledge retrieval adaptability, knowledge system redundancy ratio, effective memory decay rate, and scenario generalization and expansion capability.

[0087] Among them, the capability evolution rate represents the comprehensive improvement in the completion rate of the same type of task by the memory agent per unit time due to experience accumulation and knowledge system updates. It is used to reflect the overall efficiency of the memory agent's capability improvement, and the calculation formula can be expressed as: ; in, This represents the evolutionary dimension quantization value corresponding to the rate of capability evolution, ranging from 0 to 100. This represents the average completion rate of similar tasks at time granularity node t2. This represents the average completion rate of similar tasks at time granularity node t1. This represents the contribution weight to knowledge system updates, calculated by the number of knowledge nodes updated in the memory knowledge graph and the application effectiveness. ; The weight contributing to experience accumulation is calculated based on the reuse rate and adaptation success rate of experience nodes in the memory knowledge graph. ; For the purpose of evaluating the time period.

[0088] The knowledge retrieval suitability represents the combined value of the actual retrieval rate and retrieval efficiency of historical memory data by the memory agent in new tasks or scenarios. It reflects the actual application effect and scenario adaptability of the knowledge. The calculation formula can be expressed as: ; in, This indicates the suitability of recalling memorized knowledge, with a value ranging from 0 to 100. This represents the actual number of times a node in the application layer of the knowledge graph is invoked. This represents the total number of knowledge nodes required for a new task or scenario. This represents the number of times knowledge nodes in the application layer of the knowledge graph were successfully accessed. The weighting of the call percentage is determined by the application scenario of the memory agent. .

[0089] The redundancy ratio of a knowledge system represents the proportion of knowledge nodes in a memory knowledge graph that have no practical application value, contain repetitive content, or have logical conflicts out of the total number of knowledge nodes. It also considers the association weights and time decay of redundant nodes, reflecting the structure and effectiveness of the knowledge system. The calculation formula can be expressed as: ; in, This represents the redundancy percentage of knowledge nodes, ranging from 0 to 100. A lower value indicates better performance in memorizing the knowledge graph. This represents the number of invalid knowledge nodes that have not been reused. This represents the node importance weight of the i-th invalid knowledge node. This indicates the number of knowledge nodes with repeated content. This represents the node importance weight of the j-th repeating node. The number of knowledge nodes indicating logical conflicts. This represents the node importance weight of the k-th conflicting node. This represents the total number of knowledge nodes in the memory knowledge graph. This represents the node importance weight of the nth total knowledge node.

[0090] The effective memory decay rate represents the proportion of historically valid knowledge data that a memory agent does not reuse within a set time period. It considers both the application value of the memorized knowledge and the time decay effect, reflecting the memory agent's ability to retain and reuse valid knowledge. The calculation formula can be expressed as: ; in, This represents the effective memory decay rate, ranging from 0 to 100. A lower value indicates better memory retention. This indicates the number of valid memory nodes that have not been reused within a set time period. This represents the importance weight of the m-th unreused valid node. Indicates the time decay coefficient. It is determined by the type of knowledge being memorized. This represents the difference between the last reuse time and the evaluation time of the m-th knowledge node. This represents the total number of valid memory nodes in the memory knowledge graph. This represents the node importance weight of the e-th valid knowledge node.

[0091] The scenario generalization and expansion capability represents the effective ability of a memory agent to apply historical memory knowledge data across scenarios and tasks. It reflects the transfer and expansion level of the memory agent's capabilities, and the calculation formula can be expressed as: ; in, This indicates the ability to generalize and expand the scope of a scenario, with a value ranging from 0 to 100. This indicates the number of times a memory node in a memory knowledge graph has been successfully reused across different scenarios or tasks. This indicates the total number of scenarios across scenes or tasks. This represents the total number of tasks across the s-th scene. This represents the difficulty weight of the s-th cross-scene, calculated from the scene difference. .

[0092] This application's embodiments determine the node state data and node association data of knowledge nodes, and determine the quantification value of each evolution dimension. In this way, the evolution trajectory and memory agent are bound together through the memory knowledge graph, providing an analytical and data foundation for the generation of evolution trajectories and the location of mutation events through evolution trajectories, thereby enhancing the observation dimension of evolution trajectories and improving their comprehensiveness.

[0093] In some embodiments, after obtaining the evolution trajectory of each feature label over time based on the temporal sequence of time granularity nodes, the method further includes: Monitor incremental update events of the memory knowledge graph, including new knowledge node events, attribute change events of knowledge nodes, and weight change events of directed edges; In response to incremental update events, the evolution trajectory is dynamically updated based on the incremental data of the incremental update events.

[0094] Specifically, after generating the evolutionary trajectory, the incremental update events of the memory knowledge graph are monitored to update the memory knowledge graph and the evolutionary trajectory according to the incremental update events. The incremental update events include the addition of knowledge nodes, the change of attributes of knowledge nodes, and the change of weights of directed edges.

[0095] Among them, the new node event indicates that a new knowledge node that did not exist before has been created in the memory knowledge graph; the attribute change event indicates that the attribute value of an existing knowledge node has changed, such as the total number of times a knowledge node has been called increases, the content validity mark of a knowledge node changes from valid to redundant, or the last call timestamp of a knowledge node is updated; and the weight change event indicates that the weight value of the edge connecting two knowledge nodes has changed.

[0096] In addition, in response to the actions performed by the incremental update event, the evolutionary trajectory is dynamically updated based on the incremental data of the incremental update event. The dynamic update method can be to update the trajectory segments of the local evolutionary trajectory affected by the incremental data, that is, to recalculate and replace the feature labels corresponding to the time intervals directly related to the knowledge nodes or directed edges where the incremental update event occurred, and replace the replaced feature labels in the original evolutionary trajectory, which can improve resource utilization.

[0097] Furthermore, dynamic updates can also be performed by updating the evolutionary trajectory based on the full dataset of the memory knowledge graph.

[0098] Furthermore, the updated evolutionary trajectory can be smoothed. The smoothing process can be represented as follows: ;in, This represents the smoothed evolution trajectory value. This represents the comprehensive evolution score for the t-th evaluation period. This represents the comprehensive evolution score for the (t-1)th evaluation period. Represents the trajectory smoothing coefficient, 0 < <1, used to adjust the degree of influence of the current comprehensive evolution score and historical trajectory on the evolution trajectory value.

[0099] In this embodiment of the application, when the knowledge graph is incrementally updated, the evolutionary trajectory is refreshed accordingly, realizing the automatic and real-time refresh of the evolutionary trajectory of long-term capabilities, thereby ensuring the real-time nature and accuracy of the evaluation results. At the same time, the incremental update method improves resource utilization and update response speed.

[0100] In some embodiments, optimizing a memory agent based on mutation events includes: Determine the mutation time of the mutation event, and define the memory knowledge graph of the mutation time as the mutation knowledge graph; Based on the mutation knowledge graph, determine the capability contribution value and capability contribution direction of each knowledge node and directed edge in the mutation knowledge graph; Based on the ability contribution value and ability contribution direction, the memory knowledge graph is optimized, and the memory agent is optimized based on the optimized memory knowledge graph.

[0101] Specifically, the mutation time of the mutation event is determined, and the memory knowledge graph corresponding to the mutation event is defined as the mutation knowledge graph. A mutation event is defined as the difference between the comprehensive evolution scores of two adjacent time-granularity nodes exceeding a preset change threshold, which can be represented as: ,in, This indicates the difference in overall evolutionary scores. This indicates a preset threshold for change. This indicates a mutation event.

[0102] In addition, the capability contribution value and capability contribution direction of each knowledge node and each directed edge in the mutation knowledge graph are determined. The capability contribution value represents the magnitude of the contribution made by the knowledge node or directed edge to the change in the comprehensive evolution score at the mutation time relative to the previous moment. The capability contribution direction indicates whether the contribution value is a positive contribution that causes the comprehensive evolution score to increase or a negative contribution that causes the comprehensive evolution score to decrease.

[0103] Furthermore, based on the capability contribution value and capability contribution direction of each knowledge node and each directed edge, optimization operations are performed on the memory knowledge graph to obtain an optimized memory knowledge graph. The memory agent is then optimized based on the optimized memory knowledge graph. The optimization operations may include retaining and strengthening knowledge nodes and directed edges with positive contribution directions, and correcting or weakening knowledge nodes and directed edges with negative contribution directions.

[0104] Furthermore, the graph structure of the memory knowledge graph can undergo self-optimization, and the optimization objective can be expressed as: ; in, It represents the complete topological structure space of the memory knowledge graph. This represents the sum of the weights of all redundant knowledge nodes within the memory knowledge graph. This represents the sum of the weights of all conflicting edges within the knowledge graph. This represents the sum of weights of all valid reusable associations within the memory knowledge graph, thereby achieving the triple optimization goals of compressing invalid redundancy, resolving logical conflicts, and amplifying the value of valid reuse, thus completing the adaptive iterative optimization of the topological structure of the memory knowledge graph.

[0105] This application embodiment uses joint analysis of evolutionary trajectory and memory knowledge graph to locate the bottlenecks in the long-term capability evolution of memory agents and the underlying reasons for repeated errors, thereby obtaining targeted optimization methods for the memory knowledge graph. By optimizing the memory knowledge graph, the optimization methods for the evolution of memory agents can be determined, thereby improving the evolutionary capabilities of memory agents.

[0106] In some embodiments, based on a mutation knowledge graph, determining the capability contribution value and capability contribution direction corresponding to each knowledge node and directed edge in the mutation knowledge graph includes: Determine the trajectory deviation between the evolutionary trajectory value of the mutation time and the evolutionary trajectory value of the target capability; The trajectory deviation value is decomposed into each knowledge node and each directed edge, and the corresponding capability contribution value is determined based on the attribution ratio of each knowledge node and each directed edge to the trajectory deviation value. When there is a negative correlation between knowledge nodes or directed edges and evolution trajectory values, the direction of capability contribution is determined to be negative. When there is a positive correlation between knowledge nodes or directed edges and evolution trajectory values, the direction of capability contribution is determined to be positive.

[0107] Specifically, the evolutionary trajectory value and the target capability evolutionary trajectory value corresponding to the mutation time are determined. The evolutionary trajectory value represents the comprehensive evolutionary score corresponding to the mutation time, and the target capability evolutionary trajectory value represents the ideal capability benchmark corresponding to the mutation time.

[0108] The difference between the evolutionary trajectory value and the target capability evolutionary trajectory value is determined as the trajectory deviation value. The corresponding capability contribution value is determined according to the attribution ratio of each trajectory deviation value. The attribution ratio of a single element in the trajectory deviation value is positively correlated with its capability contribution value. That is, the higher the attribution ratio, the stronger the driving effect of the corresponding element on the current trajectory deviation.

[0109] The direction of capability contribution can be verified by calculating the marginal partial derivative. If the weight increase of a knowledge node or directed edge changes in the same direction as the increase of the evolution trajectory value, it is determined to be positively correlated, and the corresponding capability contribution direction is positive. If the weight increase of an element changes in the same direction as the decrease of the evolution trajectory value, it is determined to be negatively correlated, and the corresponding capability contribution direction is negative.

[0110] In some embodiments, the trajectory deviation value can also be decomposed into knowledge nodes and directed edges using the following formula: ; in, This represents the overall evolution score at time t. This represents the importance of the i-th knowledge node in the memory knowledge graph. This represents the state quantization value of the i-th knowledge node at time t.

[0111] ; in, This represents the weight parameter of the j-th element in the knowledge graph. The element includes knowledge nodes and directed edges. This represents the marginal contribution coefficient of the element, which serves as the attribution weight in the bias decomposition. Its positive or negative sign corresponds to the direction of the element's ability contribution.

[0112] The overall objective of optimization can be determined using a trajectory-map coupling optimization function, which can be expressed as: ; in, Represents the evolutionary trajectory value. This represents the trajectory value of the target capability evolution. This represents the graph redundancy penalty for knowledge graphs. This represents the memory decay penalty of a memory knowledge graph. and These represent the corresponding regularization weight coefficients, used to balance the priority of trajectory fit and the health of the memory knowledge graph.

[0113] The loss conclusion can be transformed into actual graph operations through a gradient-guided optimization strategy, and its calculation formula is expressed as follows: ;in, This represents the gradient vector of the loss function in the graph parameter space of the knowledge graph. The vector representing the adjustment amount of an element corresponds to outputting at least one of the following operations: deleting redundant nodes, strengthening critical paths, improving the reuse of high-value knowledge, and suppressing the recurrence of erroneous nodes.

[0114] By supplementing the attribution results with causal verification through path mining, the causal link can be represented as follows: ;in, Represents a single knowledge node. Indicates the nodes of the decision path. This indicates the execution result of the corresponding task.

[0115] The probability of a single path occurring is the product of the association probabilities of all directed edges on the causal link, and its calculation formula can be expressed as: ,in, This represents the association confidence of a single directed edge, which allows for the selection of elements with complete causal links through causal path mining, the elimination of accompanying events that occur at the same time but have no actual causal relationship, further improving the accuracy of attribution results and reducing the misadjustment of irrelevant knowledge nodes.

[0116] This application's embodiments generate targeted and implementable strategies for optimizing memory knowledge graphs through joint analysis of evolutionary trajectories and memory knowledge graphs. This provides a clear decision-making basis for the iterative optimization of memory agents, improves the optimization efficiency of memory agents, and enhances the efficiency and accuracy of anomaly handling.

[0117] In some embodiments, after optimizing the memory agent based on mutation events, the method further includes: Acquire historical time-series data of evolutionary trajectories and graph structure data of optimized memory knowledge graphs; The comprehensive evolution score of the optimized memory knowledge graph at the time of optimization completion is determined as the optimization evolution score; Based on historical time series data, graph structure data, and optimized evolution scores, predict the comprehensive future evolution score of the memory agent at at least one preset future time point.

[0118] Specifically, the historical time-series data of the evolutionary trajectory and the graph structure data of the optimized memory knowledge graph are obtained. The graph structure data includes, but is not limited to, the node importance of each knowledge node after optimization, the updated weight values ​​of each directed edge, the statistical values ​​of the calling and reuse characteristics of each knowledge node, and the topological connection relationships between knowledge nodes.

[0119] Furthermore, the comprehensive evolution score of the optimized memory knowledge graph at the optimization completion time is determined as the optimization evolution score. The optimization completion time is the point in time when all optimization operations of the memory knowledge graph have been completed and the optimized memory knowledge graph has been generated and saved. The optimization evolution score can be obtained by extracting node state data and node association data from the current state of the memory knowledge graph and calculating them.

[0120] Furthermore, based on historical time-series data, graph structure data, and optimized evolutionary scores, the comprehensive future evolutionary score of the memory agent at at least one predetermined future time point can be predicted using the following formula: ; in, This represents the comprehensive future evolution score at a predetermined future time point (t+1). This indicates an optimized evolution score. Represents historical time-series data. This represents the spectral structure data.

[0121] In some embodiments, an integrated visual display module is also included to display the memory knowledge graph, evolutionary trajectory and analysis results, which can be configured to be based on an information entropy-driven mechanism and an anomaly-priority display mechanism.

[0122] In the information entropy-driven display mechanism, the formula for calculating the information value of a knowledge node can be expressed as: ;in, Information entropy represents the information content and value of a single knowledge node. This represents the probability of calling the knowledge node in the i-th business scenario, that is, the proportion of the number of times the knowledge node is called in the i-th scenario to the total number of times the knowledge node is called.

[0123] Information entropy can be used as a filtering threshold to only display high-information nodes with information entropy higher than the preset threshold, while filtering out low-information-density edge nodes and reducing the proportion of invalid information in the visualization.

[0124] In the anomaly priority visualization mechanism, the formula for calculating the display priority can be expressed as: ;in, Indicates the display priority of map elements. Indicates the volatility weighting coefficient. This represents the influence weighting coefficient, fluctuation represents the quantitative fluctuation range of the element, and influence represents the absolute value of the element's ability contribution.

[0125] The evolutionary trajectory and memory knowledge graph are linked and compressed, and the mapping relationship can be defined as follows: ;in, This indicates a compressed mapping relationship between evolutionary trajectories and memory knowledge graphs. The structure space representing the knowledge graph includes full-dimensional information on the attributes of all knowledge nodes, the weights of directed edges, and the topological connections. It represents the feature space of the evolutionary trajectory, including the feature nodes in the evolutionary trajectory and the set of quantized values ​​of the evolutionary dimension.

[0126] This application provides an integrated, visual, and interactive display method for memory knowledge graphs and evolutionary trajectories, enabling seamless linkage and bidirectional tracing. It supports multi-dimensional and multi-time-granularity evolutionary dimension quantification values, comprehensive evolutionary score display, and personalized interactive operations, reducing the usage threshold of evolutionary trajectories and improving versatility. At the same time, it refines and personalizes the weights of evolutionary dimension quantification values ​​and predicts the evolutionary trends of memory agents' capabilities.

[0127] Figure 2 This is a schematic diagram of the architecture of a long-term capability evolution trajectory assessment system provided in an embodiment of this application, such as... Figure 2 As shown, it includes: The system comprises several modules: a full-scale knowledge acquisition module (210), a knowledge graph construction module (220), a capability evolution dimension quantification module (230), a capability evolution trajectory generation module (240), a trajectory-graph joint analysis and optimization guidance module (250), and an integrated visualization module (260). This long-term capability evolution trajectory assessment system is applied to applications such as... Figure 1 The optimization method based on long-term capability evolution trajectory shown in the figure achieves full-process, traceable, and coupled evaluation and analysis of the long-term capability evolution trajectory of the memory agent through the collaborative linkage and deep data interoperability between modules. The memory knowledge graph is the core data foundation and analysis carrier of the long-term capability evolution trajectory evaluation system. All evaluation links are carried out based on the memory knowledge graph. From the perspective of architecture design, it avoids the technical risks of evaluation bypassing memory data processing. It upgrades the evaluation dimension from static single snapshot to dynamic multi-dimensional coupled, focusing on the core elements of the agent's long-term capability evolution. It realizes the full-link systematic evaluation of memory knowledge accumulation, experience reuse, knowledge system update, and error iteration correction. Moreover, all evaluations are deeply coupled with the memory knowledge graph, which can more comprehensively, realistically, and interpretably reflect the dynamic evolution process of the agent's capability from quantitative to qualitative change. It achieves deep structured analysis and multi-dimensional correlation modeling of memory knowledge data. By constructing a three-level memory knowledge graph, it explores the inherent logical connections and causal relationships among the historical memory knowledge of memory agents, solving the problem that traditional linear storage cannot support capability evolution analysis. At the same time, it uses the memory knowledge graph as the core carrier of the system, and establishes a strong binding mechanism in terms of architecture, computing logic and data source, reducing the technical risks of bypassing memory data processing in evaluation, and providing a solid underlying data foundation for the source analysis of the capability evolution of memory agents. A quantitative evaluation system and complex computational model for the long-term capability evolution of memory agents were established. Dimensions strongly correlated with capability evolution, such as capability evolution rate, memory knowledge retrieval adaptability, and effective memory decay rate, were incorporated into the evaluation system. A complex mathematical computational model based on multiple parameters was designed to realize the quantitative characterization of the capability evolution state of memory agents and support dynamic weight adjustment. The evaluation results are more scientific, targeted, and instructive. A quantitative correlation law and two-way tracing mechanism for the accumulation of memory knowledge and the improvement of ability evolution were constructed. The mutation events of the evolution trajectory were accurately bound to specific events in the memory knowledge graph. This enabled the root cause tracing of the ability evolution trajectory and the ability impact analysis of events in the memory knowledge graph. It solved the problem of the disconnect between the existing technology assessment results and memory data, making the reasons for ability changes explainable and traceable. By generating targeted and implementable strategies for optimizing the memory knowledge graph through the joint analysis module of evolutionary trajectory and memory knowledge graph, a clear decision-making basis is provided for the iterative optimization of memory agents, which greatly improves the optimization efficiency of memory agents. In contrast, existing technologies can only output a single static evaluation result, cannot provide root cause analysis and optimization guidance, and have no closed-loop optimization mechanism. With high adaptability and scalability, this system adopts a microservice architecture and a distributed cloud storage architecture, supporting the access of different types of intelligent agents and industrial control intelligent agents in different application scenarios. It can realize the full life cycle evaluation of single-memory intelligent agents and the comparative evaluation of the capability evolution of multiple intelligent agents. At the same time, it supports the fine adjustment of evaluation dimension weights and the free configuration of analysis time granularity to meet the actual needs of different industries and different businesses. It achieves integrated visualization and two-way traceability interaction of evaluation results. By building a seamless interface linking memory knowledge graph and capability evolution trajectory, it presents complex graph data and trajectory changes in an intuitive and interactive form, supports multi-dimensional retrieval, root cause tracing, and personalized operation, lowers the threshold for using evaluation results, and is more suitable for practical engineering applications and industrialization. A combined prediction model was constructed, which integrates time series prediction and machine learning regression algorithms. Based on historical trajectory data and the update trend of memory knowledge graph, it can accurately predict the evolution trend of agent capabilities. At the same time, it can predict the effect of graph optimization on capability improvement, providing a scientific basis for the long-term iterative planning of memory agents. Existing technologies are unable to predict evolution trends.

[0128] Figure 3 This is a schematic diagram of an optimization device structure based on a long-term capability evolution trajectory provided in an embodiment of this application. Figure 3 As shown, the optimization device based on long-term capability evolution trajectory may include a construction module 310, a processing module 320, a feature module 330, an association module 340, a mutation module 350, and an optimization module 360.

[0129] The construction module 310 is used to determine the memory knowledge data of the memory agent, and to perform structured processing on the memory knowledge data according to the preset graph construction rules to obtain the memory knowledge graph corresponding to the memory agent. The processing module 320 is used to determine the evolution dimension quantization value corresponding to the memory agent based on the memory knowledge graph, and to determine the comprehensive evolution score based on the evolution dimension quantization value. The feature module 330 is used to determine the evolution dimension quantification value, comprehensive evolution score and memory knowledge graph update information corresponding to each time granularity node as the feature label of each time granularity node according to the preset time granularity. The association module 340 is used to obtain the evolution trajectory of each feature label over time based on the temporal sequence of time granularity nodes, and to establish the source association relationship between the evolution trajectory and the memory knowledge graph. The mutation module 350 is used to identify the corresponding time granularity node as a mutation node when the change in the comprehensive evolution score between adjacent time granularity nodes is greater than a preset change threshold. The optimization module 360 ​​is used to determine the mutation events corresponding to mutation nodes on the memory knowledge graph based on the source-tracing relationship, and optimize the memory agent based on the mutation events.

[0130] In some embodiments, the construction module 310 performs structured processing on the memory knowledge data according to preset graph construction rules to obtain a memory knowledge graph corresponding to the memory agent, for use in: Acquire historical memory data of the memory agent at various historical time points. The historical memory data includes historical log data, cross-scenario task association records, decision path distribution data, and task execution result data. Based on historical memory data, state difference data for each historical time node is determined. The state difference data is used to represent the changes in knowledge state, experience structure, and decision path distribution of the memory agent at adjacent time nodes. Based on the state difference data corresponding to each historical time point, multiple knowledge nodes of the memory knowledge graph are generated. Based on historical memory data, the calling relationship, reuse relationship and causal relationship between each knowledge node are determined, and multiple directed edges connecting each knowledge node are generated to obtain the memory knowledge graph corresponding to the memory agent.

[0131] In some embodiments, the construction module 310 determines the calling relationships, reuse relationships, and causal relationships between knowledge nodes based on historical memory data, generates multiple directed edges connecting each knowledge node, and obtains a memory knowledge graph corresponding to the memory agent, used for: The importance of each knowledge node is determined based on the total number of times each knowledge node is invoked in the corresponding historical log data, the total number of times it is reused across scenarios in the cross-scenario task association records, and the centrality weight value in the causal path corresponding to each causal relationship. The initial weight of each directed edge is determined based on the adjacent call time interval and call frequency between each knowledge node; Based on the interval between the last call time and the first preset time between each knowledge node, the initial weight is subjected to time decay processing to obtain decay weight; Based on the reuse frequency between knowledge nodes and the node importance of the corresponding knowledge nodes, the decay weights are reused and strengthened to obtain a dynamically evolving memory knowledge graph.

[0132] In some embodiments, the processing module 320 determines the quantization value of the evolutionary dimension corresponding to the memory agent based on the memory knowledge graph, for the purpose of: Based on the memory knowledge graph, node state data and node association data of the knowledge nodes corresponding to the memory agent are extracted. The node state data is used to represent the calling characteristics, reuse characteristics, temporal decay characteristics and content validity characteristics of each knowledge node, and the node association data is used to represent the association between each knowledge node and the task execution result and the association between each knowledge node and the usage scenario. Based on node status data and node association data, determine the quantitative value of each evolution dimension. The evolution dimension includes at least one of the following: capability evolution rate, memory knowledge retrieval adaptability, knowledge system redundancy ratio, effective memory decay rate, and scenario generalization and expansion capability.

[0133] In some embodiments, after obtaining the evolution trajectory of each feature label over time based on the temporal sequence of time granularity nodes, the association module 340 is further configured to: Monitor incremental update events of the memory knowledge graph, including new knowledge node events, attribute change events of knowledge nodes, and weight change events of directed edges; In response to incremental update events, the evolution trajectory is dynamically updated based on the incremental data of the incremental update events.

[0134] In some embodiments, the optimization module 360 ​​optimizes the memory agent based on mutation events, for the purpose of: Determine the mutation time of the mutation event, and define the memory knowledge graph of the mutation time as the mutation knowledge graph; Based on the mutation knowledge graph, determine the capability contribution value and capability contribution direction of each knowledge node and directed edge in the mutation knowledge graph; Based on the ability contribution value and ability contribution direction, the memory knowledge graph is optimized, and the memory agent is optimized based on the optimized memory knowledge graph.

[0135] In some embodiments, the optimization module 360, based on the mutation knowledge graph, determines the capability contribution value and capability contribution direction corresponding to each knowledge node and directed edge in the mutation knowledge graph, for the purpose of: Determine the trajectory deviation between the evolutionary trajectory value of the mutation time and the evolutionary trajectory value of the target capability; The trajectory deviation value is decomposed into each knowledge node and each directed edge, and the corresponding capability contribution value is determined based on the attribution ratio of each knowledge node and each directed edge to the trajectory deviation value. When there is a negative correlation between knowledge nodes or directed edges and evolution trajectory values, the direction of capability contribution is determined to be negative. When there is a positive correlation between knowledge nodes or directed edges and evolution trajectory values, the direction of capability contribution is determined to be positive.

[0136] In some embodiments, after optimizing the memory agent based on mutation events, the optimization module 360 ​​is further configured to: Acquire historical time-series data of evolutionary trajectories and graph structure data of optimized memory knowledge graphs; The comprehensive evolution score of the optimized memory knowledge graph at the time of optimization completion is determined as the optimization evolution score; Based on historical time series data, graph structure data, and optimized evolution scores, predict the comprehensive future evolution score of the memory agent at at least one preset future time point.

[0137] Figure 4 A schematic diagram of the hardware structure of the terminal device provided in an embodiment of this application is shown.

[0138] The terminal device may include a processor 401 and a memory 402 storing computer program instructions.

[0139] Specifically, the processor 401 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0140] Memory 402 may include mass storage for data or instructions. For example, and not limitingly, memory 402 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 402 may include removable or non-removable (or fixed) media, or memory 402 may be non-volatile solid-state memory. Memory 402 may be internal or external to the integrated gateway disaster recovery device.

[0141] In one instance, memory 402 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0142] Memory 402 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0143] The processor 401 reads and executes computer program instructions stored in the memory 402 to achieve... Figure 1 The optimization method based on long-term capability evolution trajectory in the illustrated embodiment.

[0144] In one example, the terminal device may also include a communication interface 403 and a bus 404. Wherein, as... Figure 4 As shown, the processor 401, memory 402, and communication interface 403 are connected through bus 404 and complete communication with each other.

[0145] The communication interface 403 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0146] Bus 404 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 404 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0147] Furthermore, in conjunction with the optimization methods based on long-term capability evolution trajectories in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the optimization methods based on long-term capability evolution trajectories in the above embodiments.

[0148] This application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the optimization methods based on long-term capability evolution trajectories described in the above embodiments.

[0149] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0150] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0151] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0152] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0153] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. An optimization method based on long-term capability evolution trajectory, characterized in that, include: The memory knowledge data of the memory agent is determined, and the memory knowledge data is structured according to the preset graph construction rules to obtain the memory knowledge graph corresponding to the memory agent. Based on the memory knowledge graph, the evolution dimension quantification value corresponding to the memory agent is determined, and the comprehensive evolution score is determined based on the evolution dimension quantification value. According to the preset time granularity, the evolution dimension quantization value, the comprehensive evolution score, and the update information of the memory knowledge graph corresponding to each time granularity node are determined as the feature labels of each time granularity node; Based on the temporal sequence of the time granularity nodes, the evolution trajectory of each feature label over time is obtained, and the source association relationship between the evolution trajectory and the memory knowledge graph is established. If the change in the comprehensive evolution score between adjacent time granularity nodes is greater than a preset change threshold, the corresponding time granularity node will be identified as a mutation node. Based on the source-tracing relationship, the mutation event corresponding to the mutation node is determined on the memory knowledge graph, and the memory agent is optimized based on the mutation event.

2. The method according to claim 1, characterized in that, The step of structuring the memory knowledge data according to preset graph construction rules to obtain the memory knowledge graph corresponding to the memory agent includes: The historical memory data of the memory agent at each historical time point is obtained, wherein the historical memory data includes historical log data, cross-scene task association records, decision path distribution data, and task execution result data; Based on the historical memory data, state difference data for each historical time node is determined, wherein the state difference data is used to represent the changes in the knowledge state, experience structure, and decision path distribution of the memory agent at adjacent time nodes; Based on the state difference data corresponding to each of the historical time nodes, multiple knowledge nodes of the memory knowledge graph are generated; Based on the historical memory data, the calling relationship, reuse relationship and causal relationship between each knowledge node are determined, and multiple directed edges connecting each knowledge node are generated to obtain the memory knowledge graph corresponding to the memory agent.

3. The method according to claim 2, characterized in that, The process involves determining the calling relationships, reuse relationships, and causal relationships between the knowledge nodes based on the historical memory data, generating multiple directed edges connecting the knowledge nodes, and obtaining the memory knowledge graph corresponding to the memory agent, including: The importance of each knowledge node is determined based on the total number of times each knowledge node is invoked in the corresponding historical log data, the total number of times it is reused across scenarios in the cross-scenario task association record, and the centrality weight value in the causal path corresponding to each causal relationship. The initial weight of each directed edge is determined based on the adjacent call time interval and call frequency between each knowledge node. Based on the interval between the last call time and the first preset time between each knowledge node, the initial weight is subjected to time decay processing to obtain decay weight; Based on the reuse frequency among the knowledge nodes and the node importance of the corresponding knowledge nodes, the decay weights are subjected to reuse and reinforcement processing to obtain the dynamically evolving memory knowledge graph.

4. The method according to claim 1, characterized in that, The step of determining the evolution dimension quantization value corresponding to the memory agent based on the memory knowledge graph includes: Based on the memory knowledge graph, node state data and node association data of the knowledge nodes corresponding to the memory agent are extracted. The node state data is used to represent the calling characteristics, reuse characteristics, temporal decay characteristics and content validity characteristics of each knowledge node. The node association data is used to represent the association between each knowledge node and the task execution result, and the association between each knowledge node and the usage scenario. Based on the node status data and the node association data, the quantitative value of each evolution dimension is determined. The evolution dimension includes at least one of the following: capability evolution rate, memory knowledge retrieval adaptability, knowledge system redundancy ratio, effective memory decay rate, and scenario generalization and expansion capability.

5. The method according to claim 1, characterized in that, After obtaining the evolution trajectory of each feature label over time based on the time granularity nodes, the method further includes: Monitor the incremental update events of the memory knowledge graph, including the addition of a knowledge node, the change of the attribute of the knowledge node, and the change of the weight of the directed edge; In response to the incremental update event, the evolutionary trajectory is dynamically updated based on the incremental data of the incremental update event.

6. The method according to claim 1, characterized in that, The optimization of the memory agent based on the mutation event includes: Determine the mutation time when the mutation event occurs, and define the memory knowledge graph of the mutation time as the mutation knowledge graph; Based on the mutation knowledge graph, determine the capability contribution value and capability contribution direction of each knowledge node and the directed edge in the mutation knowledge graph; Based on the capability contribution value and the capability contribution direction, the memory knowledge graph is optimized, and the memory agent is optimized based on the optimized memory knowledge graph.

7. The method according to claim 6, characterized in that, The step of determining the capability contribution value and capability contribution direction corresponding to each knowledge node and directed edge in the mutation knowledge graph based on the mutation knowledge graph includes: Determine the trajectory deviation between the evolutionary trajectory value of the mutation time and the evolutionary trajectory value of the target capability; The trajectory deviation value is decomposed into each of the knowledge nodes and each of the directed edges, and the corresponding capability contribution value is determined based on the attribution ratio of each of the knowledge nodes and each of the directed edges to the trajectory deviation value. When there is a negative correlation between the knowledge node or the directed edge and the evolution trajectory value, the capability contribution direction is determined to be negative. When there is a positive correlation between the knowledge node or the directed edge and the evolution trajectory value, the capability contribution direction is determined to be positive.

8. The method according to claim 1, characterized in that, After optimizing the memory agent based on the mutation event, the method further includes: Obtain the historical time-series data of the evolutionary trajectory and the optimized graph structure data of the memory knowledge graph; The comprehensive evolution score of the optimized memory knowledge graph at the optimization completion time is determined as the optimization evolution score; Based on the historical time series data, the graph structure data, and the optimized evolution score, the comprehensive future evolution score of the memory agent at at least one preset future time point is predicted.

9. An optimization device based on a long-term capability evolution trajectory, characterized in that, The device includes: The construction module is used to determine the memory knowledge data of the memory agent, and to perform structured processing on the memory knowledge data according to the preset graph construction rules to obtain the memory knowledge graph corresponding to the memory agent. The processing module is used to determine the evolution dimension quantization value corresponding to the memory agent based on the memory knowledge graph, and to determine the comprehensive evolution score based on the evolution dimension quantization value; The feature module is used to determine the evolution dimension quantization value, the comprehensive evolution score and the update information of the memory knowledge graph corresponding to each time granularity node as the feature label of each time granularity node according to a preset time granularity. The association module is used to obtain the evolution trajectory of each feature label over time based on the temporal sequence of the time granularity nodes, and to establish the source association relationship between the evolution trajectory and the memory knowledge graph. The mutation module is used to determine the corresponding time granularity node as a mutation node when the change in the comprehensive evolution score between adjacent time granularity nodes is greater than a preset change threshold. An optimization module is used to determine the mutation event corresponding to the mutation node on the memory knowledge graph based on the source association relationship, and optimize the memory agent based on the mutation event.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the optimization method based on long-term capability evolution trajectory as described in any one of claims 1-8.