A long-term memory construction and generation method for cognitive AI hardware

Abstract

The application relates to the technical field of artificial intelligence, and discloses a long-term memory construction and generation method for cognitive AI hardware, which comprises the following steps: performing semantic understanding and multidimensional analysis on interactive information of a user to obtain multidimensional memory features; performing self-adaptive matching derivation on the multidimensional memory features and a preset long-term memory knowledge graph, integrating memory subgraphs into the long-term memory knowledge graph, and obtaining an updated long-term memory; performing double-end embedded retrieval on key factual data and the updated long-term memory to obtain a bidirectional mapping relationship; performing similarity retrieval on new round interactive information based on the bidirectional mapping relationship to obtain an associated memory segment; performing dynamic weight distribution on the associated memory segment to obtain an optimized memory segment; and performing adaptive deduction on the updated long-term memory based on the optimized memory segment to obtain personalized response content; and the application can improve the efficiency of the long-term memory construction and generation method for the cognitive AI hardware.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for constructing and generating long-term memory for cognitive AI hardware. Background Technology

[0002] Existing methods for building long-term memory in cognitive AI hardware struggle to accurately extract key factual data and transform it into effective memory features when processing multi-dimensional user interaction information, resulting in insufficient targeting and completeness in memory construction. Furthermore, the fusion of memory subgraphs and long-term memory knowledge graphs lacks an adaptive matching mechanism, making it difficult to balance structural adaptability and attribute correlation. This leads to low efficiency in updating long-term memory and an inability to quickly meet personalized user needs.

[0003] Traditional methods lack efficient dual-end embedded retrieval logic and dynamic weight allocation strategies in the memory retrieval and response generation stages. This results in insufficient accuracy in filtering associated memory fragments, leading to a low degree of alignment between personalized response content and the user's actual needs, and poor practicality and timeliness of generated reference information. Therefore, improving the efficiency of long-term memory construction and generation methods for cognitive AI hardware has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method for constructing and generating long-term memory for cognitive AI hardware, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for constructing and generating long-term memory for cognitive AI hardware, comprising: S1. Perform semantic understanding on the user's interaction information to obtain the user's key fact data, and perform multi-dimensional analysis on the key fact data to obtain the user's multi-dimensional memory features; S2. Adaptively match and deduce the multidimensional memory features with the preset long-term memory knowledge graph to obtain the user's fusion strategy, and based on the fusion strategy, integrate the user's memory subgraph into the long-term memory knowledge graph to obtain the user's updated long-term memory. S3. Perform a two-way embedded retrieval on the key fact data and the updated long-term memory to obtain the bidirectional mapping relationship of the user; S4. Based on the bidirectional mapping relationship, perform long-term memory similarity retrieval on the user's new round of interaction information to obtain the user's associated memory fragments; S5. Perform dynamic weight allocation on the associated memory fragments to obtain the user's optimized memory fragments; S6. Based on the optimized memory fragment, the updated long-term memory is adapted and deduced to obtain the user's personalized response content.

[0006] In a preferred embodiment, the semantic understanding of the user's interaction information to obtain the user's key fact data, and the multi-dimensional analysis of the key fact data to obtain the user's multi-dimensional memory features, include: The system acquires user interaction information, which includes voice stream text, dialogue sequences, image information content, and emotional state parameters. Context-aware semantic deconstruction is performed on the interaction information to obtain the user's basic fact elements; The basic fact elements are categorized and merged to obtain the user's key fact data; Perform cross-dimensional feature mapping on the key fact data to obtain the user's primary characteristics; The primary features are fused to obtain the user's multidimensional memory features.

[0007] In a preferred embodiment, the step of adaptively matching and deriving the multidimensional memory features with a preset long-term memory knowledge graph to obtain the user's fusion strategy, and integrating the user's memory subgraph into the long-term memory knowledge graph based on the fusion strategy to obtain the user's updated long-term memory, includes: The multidimensional memory features are used to construct a feature graph to obtain the user's memory subgraph; Structural topology analysis is performed on the memory subgraph and the long-term memory knowledge graph to obtain the structural matching degree of the user. Based on the structural matching degree, attribute compatibility identification is performed on the memory subgraph to obtain the attribute association degree of the user; A multi-dimensional decision derivation is performed on the structural matching degree and the attribute correlation degree to obtain the user's fusion strategy; Based on the fusion strategy, the memory subgraph is reconstructed in relational topology to obtain the user's enhanced memory subgraph; The enhanced memory subgraph and the long-term memory knowledge graph are integrated and optimized to obtain the user's updated long-term memory.

[0008] In a preferred embodiment, the step of performing multi-dimensional decision derivation on the structural matching degree and the attribute correlation degree to obtain the user's fusion strategy includes: By performing state correlation between the structural matching degree and the attribute correlation degree, the user's decision characteristics are obtained; Based on the decision characteristics, heuristic deduction is performed on the strategy generation of the memory subgraph to obtain the user's strategy derivation path; The strategy derivation path is backtracked and verified to obtain the user's optimized strategy; The optimized strategy is instantiated and adapted to obtain the user's fusion strategy.

[0009] In a preferred embodiment, the step of performing a two-way embedded retrieval of the key fact data and the updated long-term memory to obtain the bidirectional mapping relationship of the user includes: The key fact data is feature-encoded to obtain the memory feature identifier of the key fact data; Based on the memory feature identifier, the entries in the updated long-term memory are matched and identified to obtain the user's bound memory entries; The memory feature identifier is matched with the bound memory entry to obtain the index association representation of the user; Based on the index association representation, bidirectional entry writing is performed on the user's index storage area to obtain the user's bidirectional mapping relationship.

[0010] In a preferred embodiment, the step of performing long-term memory similarity retrieval on the user's new round of interaction information based on the bidirectional mapping relationship to obtain the user's associated memory fragments includes: Perform interactive semantic analysis on the user's new round of interaction information to obtain the user's query semantic features; Based on the bidirectional mapping relationship, the query semantic features are transformed into a vector space to obtain the user's query vector; Based on the query vector, a proximity search and matching is performed on the updated long-term memory to obtain the user's associated memory fragments.

[0011] In a preferred embodiment, the step of dynamically weighting the associated memory fragments to obtain the user's optimized memory fragments includes: The associated memory fragments are deconstructed using multidimensional features to obtain the feature spectrum of the associated memory fragments. The weighted influence factors are extracted from the intrinsic correlation dimension of the feature spectrum to obtain the initial influence factors of the associated memory fragments; The initial influence factor is subjected to multi-source information collaborative deduction to obtain the enhanced influence factor of the associated memory fragment; The enhanced influence factors are subjected to contextual consistency fusion to obtain the dynamic weights of the associated memory segments; Based on the dynamic weights, the information structure of the associated memory fragments is reconstructed to obtain the user's optimized memory fragments.

[0012] In a preferred embodiment, the dynamic weight is calculated using the following formula: ; in, For the dynamic weight, For the first The aforementioned enhancing influence factors, For the first The contextual consistency coefficient of each of the associated memory segments, For the first The feature dimension weights of the associated memory segments The mean of the enhanced influence factor, The mean of the contextual consistency coefficients of the associated memory segments. This is a global correction factor. The current interaction time. For the first The generation time of the aforementioned associated memory fragments, The time decay coefficient, This represents the total number of the associated memory fragments.

[0013] In a preferred embodiment, the step of identifying and extracting weighted influence factors for the intrinsic correlation dimension of the feature spectrum to obtain the initial influence factors of the associated memory fragment includes: The associated memory fragments are metrically analyzed to obtain the activity and significance metrics of the associated memory fragments; Cross-validation is performed on the activity measure and the correlation measure to obtain the validated correlation factor of the associated memory fragment; The importance of the verified association factors is ranked to obtain the ranking factors of the associated memory segments; The sorted factors are structured and encapsulated to obtain the initial influence factors of the associated memory fragments.

[0014] In a preferred embodiment, the step of adapting and extrapolating the updated long-term memory based on the optimized memory fragment to obtain the user's personalized response content includes: The optimized memory fragments are refined to obtain the user's focus memory elements; Based on the aforementioned focal memory elements, a structured text description is extracted from the updated long-term memory to obtain the user's long-term memory text. A semantic field is constructed by combining the long-term memory text with the user's short-term context data to obtain the user's fused memory context. The fused memory context is serialized and arranged into instructions to obtain the user's generated instruction sequence; The user's personalized response content is obtained by inferring the intent from the generated instruction sequence.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention utilizes semantic understanding and multi-dimensional analysis of user interaction information to accurately extract key factual data and transform it into multi-dimensional memory features. Combined with an adaptive matching and derivation mechanism, it efficiently integrates memory subgraphs into the long-term memory knowledge graph, significantly improving the targeting and completeness of long-term memory construction, allowing long-term memory to quickly meet users' personalized needs. Simultaneously, the synergistic application of dual-end embedded retrieval and similarity retrieval greatly optimizes the accuracy and efficiency of memory retrieval, ensuring rapid location and extraction of associated memory fragments.

[0016] 2. This invention optimizes and reconstructs associated memory fragments using a dynamic weight allocation strategy, and combines this with adaptive deduction to generate personalized response content. This effectively improves the alignment between the response content and user needs, enhancing the practicality and timeliness of the reference information. The entire process forms a full-link optimization from memory construction and retrieval to response generation, comprehensively improving the construction quality and application efficiency of long-term memory in cognitive AI hardware, and providing users with a more accurate and efficient interactive experience. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for constructing and generating long-term memory for cognitive AI hardware, provided in an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a method for constructing and generating long-term memory for cognitive AI hardware. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for constructing and generating long-term memory for cognitive AI hardware can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a method for constructing and generating long-term memory for cognitive AI hardware according to an embodiment of the present invention. In this embodiment, the method for constructing and generating long-term memory for cognitive AI hardware includes: S1. Perform semantic understanding on the user's interaction information to obtain the user's key fact data, and perform multi-dimensional analysis on the key fact data to obtain the user's multi-dimensional memory features; In this embodiment of the invention, the step of performing semantic understanding on the user's interaction information to obtain the user's key fact data, and performing multi-dimensional analysis on the key fact data to obtain the user's multi-dimensional memory features, includes: The system acquires user interaction information, which includes voice stream text, dialogue sequences, image information content, and emotional state parameters. Context-aware semantic deconstruction is performed on the interaction information to obtain the user's basic fact elements; The basic fact elements are categorized and merged to obtain the user's key fact data; Perform cross-dimensional feature mapping on the key fact data to obtain the user's primary characteristics; The primary features are fused to obtain the user's multidimensional memory features.

[0021] Voice signals during user interaction with cognitive AI hardware are captured in real time by voice acquisition devices and converted into coherent speech-to-text technology, ensuring that the voice information is completely and accurately transformed into a processable text format. Simultaneously, the interaction recording module records the entire dialogue between the user and the AI ​​hardware, organizing it chronologically to form a dialogue sequence that clearly presents the logical flow of the interaction. Image acquisition devices capture various image information involved in the interaction scenario, including images of objects displayed by the user and environmental images, extracting key visual elements from the images. Emotion perception sensors monitor changes in the user's voice tone and facial expressions, quantifying this emotion-related information into specific emotional state parameters. Finally, the speech stream text, dialogue sequence, image information content, and emotional state parameters are integrated and aggregated to form comprehensive and rich user interaction information.

[0022] By deeply integrating the user's current interaction scenario, past dialogue history, and relevant background knowledge, the collected interaction information undergoes meticulous semantic decomposition, word by word, sentence by sentence, and section by section. For voice stream text and dialogue sequences, the logical relationships and semantic connections between sentences are analyzed to identify core information such as the identity of the person, specific events, time of occurrence, location involved, and core needs. For image information, the semantic meaning conveyed by the images is interpreted, key visual information related to the interaction theme is extracted, and converted into factual elements in textual form. Combined with emotional state parameters, the core intent expressed by the user is determined, ensuring that no key information is missed. In this way, complex interaction information is decomposed into independent, clear, and accurate basic factual elements that reflect the core content of the interaction.

[0023] All the basic factual elements obtained from the decomposition are systematically classified and organized, establishing clear classification standards according to information type. For example, all basic factual elements involving specific times are grouped into the time category, those involving specific events into the event category, those involving user needs into the need category, and those involving people into the people category, etc. During the classification process, duplicate and redundant basic factual elements are carefully identified, and content that is completely identical or has the same core information is eliminated to avoid information redundancy. At the same time, the basic factual elements under each category are further filtered and refined, retaining information that can comprehensively and accurately reflect the user's core intent and key situations. This classified, deduplicated, and filtered information is then integrated to ultimately form structured and organized user key factual data.

[0024] For different types of information in key factual data, a multi-dimensional feature mapping framework is constructed to clarify the feature dimensions corresponding to each information type. For example, time-related key factual data is mapped to a time feature dimension, specifically including features such as time points and time spans; demand-related key factual data is mapped to a demand feature dimension, covering features such as the type of demand, urgency, and specific requirements; and person-related key factual data is mapped to a person feature dimension, including features such as the person's identity, preferences, and relationships. Following this pre-defined multi-dimensional feature mapping framework, each type of key factual data is transformed into a corresponding feature representation, ensuring that each key factual data can find its corresponding position in the feature mapping framework, thereby obtaining comprehensive user primary features covering all aspects of the key factual data.

[0025] A feature concatenation approach is used to integrate primary features from various dimensions. A pre-defined concatenation order ensures the standardization and orderliness of the process. During concatenation, the original information of each primary feature is strictly preserved to prevent loss or distortion, fully retaining the specific feature content of each dimension. Through this orderly concatenation, primary features scattered across different dimensions are combined into a complete and unified feature set. This feature set comprehensively covers all core information of key factual data, reflecting both the independence of each dimension's features and the organic integration of feature information, ultimately forming a multi-dimensional memory feature that accurately reflects the user's key information.

[0026] The beneficial effects are that, through a step-by-step and refined implementation process, user interaction information is comprehensively collected from multiple channels. After a series of standardized operations such as semantic deconstruction, category merging, feature mapping and fusion, key factual data is accurately extracted and transformed into multi-dimensional memory features. This not only solves the problems of inaccurate extraction of key factual data and incomplete transformation of memory features in existing methods, but also ensures that the products generated in each step are accurate and complete. This provides high-quality and reliable basic data support for subsequent memory subgraph construction, long-term memory updates and personalized response generation, significantly improving the targeting and effectiveness of long-term memory construction in cognitive AI hardware.

[0027] S2. Adaptively match and deduce the multidimensional memory features with the preset long-term memory knowledge graph to obtain the user's fusion strategy, and based on the fusion strategy, integrate the user's memory subgraph into the long-term memory knowledge graph to obtain the user's updated long-term memory. In this embodiment of the invention, the adaptive matching and derivation of the multidimensional memory features and the preset long-term memory knowledge graph to obtain the user's fusion strategy, and the integration of the user's memory subgraph into the long-term memory knowledge graph based on the fusion strategy to obtain the user's updated long-term memory, includes: The multidimensional memory features are used to construct a feature graph to obtain the user's memory subgraph; Structural topology analysis is performed on the memory subgraph and the long-term memory knowledge graph to obtain the structural matching degree of the user. Based on the structural matching degree, attribute compatibility identification is performed on the memory subgraph to obtain the attribute association degree of the user; A multi-dimensional decision derivation is performed on the structural matching degree and the attribute correlation degree to obtain the user's fusion strategy; Based on the fusion strategy, the memory subgraph is reconstructed in relational topology to obtain the user's enhanced memory subgraph; The enhanced memory subgraph and the long-term memory knowledge graph are integrated and optimized to obtain the user's updated long-term memory.

[0028] The process of performing multi-dimensional decision derivation on the structural matching degree and the attribute correlation degree to obtain the user's fusion strategy includes: By performing state correlation between the structural matching degree and the attribute correlation degree, the user's decision characteristics are obtained; Based on the decision characteristics, heuristic deduction is performed on the strategy generation of the memory subgraph to obtain the user's strategy derivation path; The strategy derivation path is backtracked and verified to obtain the user's optimized strategy; The optimized strategy is instantiated and adapted to obtain the user's fusion strategy.

[0029] By sorting out the relationships between the features in the multidimensional memory feature, and constructing a graph structure with nodes and edges according to the logical connections and hierarchical structure between the features, each feature is treated as an independent node, and the relationships between features are treated as edges connecting the nodes, forming a user memory subgraph that can fully present the internal relationships of the multidimensional memory features.

[0030] The memory subgraph is compared and analyzed with the pre-defined long-term memory knowledge graph. The structural elements such as the number of nodes, hierarchical distribution, and connection methods of the two are compared one by one. The overlapping parts and differences between the memory subgraph and the long-term memory knowledge graph in terms of structure are counted. The structural matching degree between the two is determined based on the proportion of the overlapping parts. The structural matching degree can intuitively reflect the degree of fit between the memory subgraph and the long-term memory knowledge graph at the structural level.

[0031] Based on structural matching, this study delves into the attribute information of each node in the memory subgraph, including attribute type, attribute value, and attribute constraints. This attribute information is compared with the attribute information of corresponding nodes in the long-term memory knowledge graph to determine whether the attributes are compatible, conflicting, or complementary. Based on the degree and scope of attribute compatibility, the attribute correlation between the memory subgraph and the long-term memory knowledge graph is determined. The numerical information corresponding to the structural matching and attribute correlation are then integrated and analyzed as a whole, extracting core information reflecting their collaborative relationship to form user decision characteristics. These decision characteristics centrally reflect the comprehensive adaptation of the memory subgraph and the long-term memory knowledge graph at both the structural and attribute levels.

[0032] Based on decision-making characteristics, combined with the construction rules of long-term memory knowledge graphs and the goals of memory integration, starting from the existing structure and attributes of memory subgraphs, we gradually deduce possible fusion directions and methods, explore the feasibility and effects of different fusion paths, and form multiple potential strategy derivation paths, each path corresponding to a possible fusion scheme.

[0033] For each strategy derivation path, reverse trace is performed to verify the rationality and accuracy of each derivation step in the path, check for logical loopholes, conflicts with preset rules, or situations that cannot be implemented, eliminate problematic paths, and retain the strategies corresponding to logically rigorous and feasible paths to obtain the optimized strategy.

[0034] Based on the application scenarios of cognitive AI hardware, the storage characteristics of long-term memory, and the actual needs of user interaction, the optimized strategy is specifically adapted for implementation. Details such as the operational steps, execution order, and parameter settings involved in the strategy are clarified, transforming the abstract strategy into a directly executable concrete solution, ultimately forming the user's fusion strategy. Following the rules and requirements specified in the fusion strategy, the node relationships and hierarchical structure of the memory subgraph are reorganized and reconstructed, supplementing missing relationships, optimizing unreasonable connection methods, and strengthening the connections between key nodes. This makes the structure of the memory subgraph more consistent with the integration standards of long-term memory knowledge graphs, forming a more complete and tightly connected augmented memory subgraph.

[0035] The augmented memory subgraph is fused with the long-term memory knowledge graph node by node and relation by relation. The parameters such as the node position and relation weight of the augmented memory subgraph are adjusted to make it seamlessly connected with the long-term memory knowledge graph. At the same time, the information redundancy and conflict that may occur during the fusion process are eliminated to ensure that the integrated knowledge graph has a unified structure and consistent information, and finally the updated long-term memory is obtained.

[0036] The beneficial effects are that by first constructing a memory subgraph, then conducting dual adaptation analysis at the structural and attribute levels, and combining multi-dimensional decision deduction to determine a scientific and reasonable fusion strategy, the memory subgraph and long-term memory knowledge graph are efficiently integrated through relationship reconstruction and chimera optimization. This ensures the targeting and accuracy of the integration process, improves the update efficiency of long-term memory, and enables long-term memory to quickly absorb new memory information from users, better meeting users' personalized needs.

[0037] S3. Perform a two-way embedded retrieval on the key fact data and the updated long-term memory to obtain the bidirectional mapping relationship of the user; In this embodiment of the invention, the step of performing a two-way embedded retrieval of the key fact data and the updated long-term memory to obtain the bidirectional mapping relationship of the user includes: The key fact data is feature-encoded to obtain the memory feature identifier of the key fact data; Based on the memory feature identifier, the entries in the updated long-term memory are matched and identified to obtain the user's bound memory entries; The memory feature identifier is matched with the bound memory entry to obtain the index association representation of the user; Based on the index association representation, bidirectional entry writing is performed on the user's index storage area to obtain the user's bidirectional mapping relationship.

[0038] The key fact data is sorted out one by one, and each dimension of key fact data is transformed into a unique combination of characters or identification code according to the preset coding rules. This coding process strictly corresponds to the core content of the key fact data, ensuring that each key fact data can generate a unique and identifiable memory feature identifier, which can accurately represent the essential characteristics of the corresponding key fact data.

[0039] Using the generated memory feature identifier as the retrieval basis, all entries in the updated long-term memory are traversed, and the feature information contained in each entry is compared with the feature represented by the memory feature identifier. Entries that are completely matched or highly related to the memory feature identifier in terms of features are selected, and these selected entries are marked as bound memory entries corresponding to the memory feature identifier, ensuring that the bound memory entries have a clear correspondence with the key fact data.

[0040] Establish a rule for associating memory feature identifiers with their corresponding bound memory entries. Based on this rule, associate each memory feature identifier with all its corresponding bound memory entries one-to-one. Record the details of the association between the two, including the matching points and the strength of the association. Organize these associations into a structured index association representation, which clearly presents the correspondence logic between memory feature identifiers and bound memory entries.

[0041] Open the user's index storage area, and write the memory feature identifier and its corresponding bound memory entry as the bidirectional association entry content into the index storage area according to the association relationship recorded in the index association representation. This ensures that in the index storage area, the corresponding bound memory entry can be quickly queried through the memory feature identifier, and the corresponding memory feature identifier can also be traced back through the bound memory entry, ultimately forming a complete bidirectional mapping relationship.

[0042] The beneficial effects are that through standardized feature coding, accurate entry matching, clear association representation, and bidirectional write operations, efficient dual-end embedded retrieval of key fact data and updated long-term memory is achieved. The constructed bidirectional mapping relationship provides a fast and accurate index foundation for subsequent similarity retrieval, greatly improving the efficiency and accuracy of memory retrieval, and solving the problem of low retrieval efficiency caused by unclear indexes and ambiguous associations in traditional retrieval methods.

[0043] S4. Based on the bidirectional mapping relationship, perform long-term memory similarity retrieval on the user's new round of interaction information to obtain the user's associated memory fragments; In this embodiment of the invention, the step of performing long-term memory similarity retrieval on the user's new round of interaction information based on the bidirectional mapping relationship to obtain the user's associated memory fragments includes: Perform interactive semantic analysis on the user's new round of interaction information to obtain the user's query semantic features; Based on the bidirectional mapping relationship, the query semantic features are transformed into a vector space to obtain the user's query vector; Based on the query vector, a proximity search and matching is performed on the updated long-term memory to obtain the user's associated memory fragments.

[0044] We comprehensively collect all kinds of information generated during the user's new round of interaction, including text converted from voice communication, text dialogue content, image information involved in the interaction, and related operation instructions. We conduct detailed analysis of this information sentence by sentence and part by part, and combine it with the current interaction scenario and contextual logic to deconstruct the core semantics, user intent, and key expressions contained therein. We extract the core semantic information that can reflect the user's current query needs, and organize and summarize this information to form the user's query semantic features, ensuring that the features can accurately correspond to the core needs of the user's new round of interaction.

[0045] Based on the established bidirectional mapping relationship, the corresponding association rules between query semantic features and various entries in the long-term memory knowledge graph are clarified. According to the rules, each core semantic information of the query semantic features is transformed into corresponding coordinate points in the vector space. Through a standardized conversion process, the scattered semantic information is integrated into a unified and computable vector form. This vector fully retains the core connotation and association of the query semantic features, and finally forms the user's query vector.

[0046] Using the generated query vector as the retrieval benchmark, all entries in the updated long-term memory are traversed. The distance between the query vector and the vector corresponding to each entry in the long-term memory is calculated. The closer the distance, the higher the semantic similarity between the two. All entries whose distance to the query vector is within a preset range are selected. These selected entries are further semantically matched and verified to confirm their relevance to the user's new round of interaction needs. The memory information corresponding to the verified entries is integrated to form the user's associated memory fragments.

[0047] The beneficial effects are that by first analyzing the new round of interaction information to obtain accurate query semantic features, then using bidirectional mapping to complete vector transformation, and finally achieving efficient matching through proximity retrieval, the accuracy of the screening of associated memory fragments and the retrieval efficiency are ensured. This effectively solves the problems of inaccurate semantic matching and slow retrieval speed in traditional retrieval, and provides core memory materials that meet user needs for subsequent dynamic weight allocation and personalized response generation.

[0048] S5. Perform dynamic weight allocation on the associated memory fragments to obtain the user's optimized memory fragments; In this embodiment of the invention, the step of dynamically weighting the associated memory fragments to obtain the user's optimized memory fragments includes: The associated memory fragments are deconstructed using multidimensional features to obtain the feature spectrum of the associated memory fragments. The weighted influence factors are extracted from the intrinsic correlation dimension of the feature spectrum to obtain the initial influence factors of the associated memory fragments; The initial influence factor is subjected to multi-source information collaborative deduction to obtain the enhanced influence factor of the associated memory fragment; The enhanced influence factors are subjected to contextual consistency fusion to obtain the dynamic weights of the associated memory segments; Based on the dynamic weights, the information structure of the associated memory fragments is reconstructed to obtain the user's optimized memory fragments.

[0049] The formula for calculating the dynamic weight is as follows: ; in, For the dynamic weight, For the first The aforementioned enhancing influence factors, For the first The contextual consistency coefficient of each of the associated memory segments, For the first The feature dimension weights of the associated memory segments The mean of the enhanced influence factor, The mean of the contextual consistency coefficients of the associated memory segments. This is a global correction factor. The current interaction time. For the first The generation time of the aforementioned associated memory fragments, The time decay coefficient, This represents the total number of the associated memory fragments.

[0050] The step of identifying and extracting weighted influence factors for the intrinsic correlation dimension of the feature spectrum to obtain the initial influence factors of the associated memory fragment includes: The associated memory fragments are metrically analyzed to obtain the activity and significance metrics of the associated memory fragments; Cross-validation is performed on the activity measure and the correlation measure to obtain the validated correlation factor of the associated memory fragment; The importance of the verified association factors is ranked to obtain the ranking factors of the associated memory segments; The sorted factors are structured and encapsulated to obtain the initial influence factors of the associated memory fragments.

[0051] Each associated memory fragment is comprehensively and meticulously dissected, deeply exploring its core information across multiple dimensions, including semantic features, relational features, temporal features, and user attention features. The specific manifestations and connotations of each dimension are systematically analyzed, and these features from different dimensions are organized into a clear and hierarchical feature set, forming a feature spectrum that can fully present all aspects of the associated memory fragment's attributes, ensuring that the feature spectrum comprehensively covers the key information of the associated memory fragment.

[0052] The activity level of associated memory fragments is quantitatively evaluated. The activity level is mainly determined by the frequency of retrieval of the fragment and the degree of relevance to the current interaction needs, thus obtaining an activity metric that reflects the current effectiveness of the fragment. At the same time, the importance of the information contained in the associated memory fragments to the user's current query needs is analyzed. The importance is judged based on the degree of matching between the information and the user's core needs and the uniqueness of the information, thus obtaining a significant metric that reflects the value of the fragment.

[0053] The obtained activity measure and significance measure are cross-validated to compare whether the characteristics of the associated memory fragments reflected by the two are consistent. If there are differences, the judgment criteria and basis in the measurement process are re-examined, the bias is corrected, and the two measurement results are ensured to support each other and complement each other. Through this cross-validation method, inaccurate and unreasonable measurement components are eliminated, and a verified association factor that is more in line with the actual situation of associated memory fragments is obtained.

[0054] A clear importance evaluation criterion is established. This criterion comprehensively considers factors such as the contribution of the verified correlation factors to the user's current interaction needs and their fit with the semantic features of the query. All verified correlation factors are evaluated and scored one by one according to this criterion. These factors are ranked according to their scores. The higher the score, the higher the importance of the factor and the higher the ranking. Finally, an ordered ranking of factors is formed.

[0055] Following a pre-defined structured format, the sorted factors are organized and packaged, clearly defining the name, specific meaning, importance score, and corresponding associated memory fragments of each factor. This ensures that the packaged initial impact factors have a clear structure and definite direction, enabling them to be directly used in subsequent collaborative deduction processes and guaranteeing the standardization and usability of the initial impact factors.

[0056] Collect multi-source auxiliary information related to the associated memory fragments, including the fragment's historical usage records, user feedback on similar fragments, and background information of the current interaction scenario. Combine this multi-source information with the initial influencing factor, deeply analyze the internal connections and interactions between the various information sources, deduce the changing trend and strengthening direction of the initial influencing factor under the support of multi-source information, supplement key information not covered in the initial influencing factor, improve the comprehensiveness and accuracy of the factor, and ultimately form an enhanced influencing factor.

[0057] By combining the contextual information of the current user interaction, including the core semantics of the new round of interaction information, the previous dialogue logic, and the user's current needs scenario, we analyze the degree of fit between each enhanced influence factor and the contextual information, determine whether the information represented by the factor is consistent with the context, assign a corresponding consistency weight to each enhanced influence factor according to the degree of fit, and fuse the enhanced influence factor with the corresponding consistency weight to obtain a dynamic weight that reflects the importance of the factor in the current context environment.

[0058] This calculation integrates the core influencing factors, contextual fit, and feature dimension importance of associated memory fragments, combines the overall average level with global correction adjustment, and introduces a time decay mechanism to consider the timeliness of fragments. Finally, it obtains a dynamic value that can accurately reflect the importance of each associated memory fragment, providing a scientific basis for the reconstruction of the information structure of associated memory fragments.

[0059] The enhanced impact factor is derived from the multi-dimensional feature deconstruction of associated memory fragments, the identification and extraction of initial impact factors, and the collaborative deduction of multi-source information. The context consistency coefficient is determined by the consistency fusion of the enhanced impact factor and the context of the associated memory fragment. The feature dimension weight is the inherent feature dimension related weight of the associated memory fragment itself. The mean of the enhanced impact factor is the arithmetic mean of all enhanced impact factors. The mean of the context consistency coefficient is the arithmetic mean of the context consistency coefficients of all associated memory fragments. The global correction coefficient is a preset fixed parameter, which is usually between 0.1 and 0.5 in this dynamic weight calculation scenario. The current interaction time is the specific time point when a new round of interaction information is obtained. The generation time of the associated memory fragment is the specific time point when the fragment was initially constructed. The time decay coefficient is a preset fixed parameter, which is based on the time span characteristics of the associated memory fragments and combined with the information timeliness requirements in the user interaction scenario. Through comparative experiments on the retrieval and response effects of memory fragments with multiple time intervals, the impact of dynamic weights under different coefficients on response accuracy is statistically analyzed, and finally, the fixed parameters suitable for this scenario are determined. The total number of associated memory fragments is the total number of fragments obtained after similarity retrieval.

[0060] As the weights of the enhanced impact factor, contextual consistency coefficient, or feature dimension increase, the dynamic value will increase accordingly. As the mean of the enhanced impact factor or the mean of the contextual consistency coefficient increases, the dynamic value will increase accordingly. The greater the difference between the current interaction time and the generation time of the associated memory fragment, the greater the result of the time decay part, and the dynamic value will also increase accordingly. When the total number of associated memory fragments increases, if the growth rate of the numerator is less than that of the denominator, the dynamic value will decrease accordingly. The increase of the global correction coefficient and the time decay coefficient will affect the increase of the dynamic value through correction adjustment and time decay effects, respectively.

[0061] Based on the dynamic weights, the information in the associated memory fragments is reordered, with information with higher dynamic weights placed in more prominent positions and information with lower weights placed later. At the same time, key information with higher weights is integrated, while redundant information with extremely low weights and no practical value to user needs is removed. The presentation structure and logical relationship of the information are optimized so that the reconstructed information can accurately focus on the core needs of users, forming optimized memory fragments with a reasonable structure and prominent key points.

[0062] The beneficial effects are that through multi-dimensional feature deconstruction, scientific measurement and verification, orderly sorting and encapsulation, comprehensive collaborative inference, and context-appropriate fusion calculation, the dynamic weight of associated memory fragments is accurately allocated. The information structure reconstruction based on dynamic weights makes the optimized memory fragments more in line with the user's current needs, effectively solving the problem of insufficient accuracy in memory fragment selection in traditional methods, providing high-quality core materials for subsequent personalized response generation, and improving the relevance and practicality of the response content.

[0063] S6. Based on the optimized memory fragment, the updated long-term memory is adapted and deduced to obtain the user's personalized response content.

[0064] In this embodiment of the invention, the step of adapting and extrapolating the updated long-term memory based on the optimized memory fragment to obtain the user's personalized response content includes: The optimized memory fragments are refined to obtain the user's focus memory elements; Based on the aforementioned focal memory elements, a structured text description is extracted from the updated long-term memory to obtain the user's long-term memory text. A semantic field is constructed by combining the long-term memory text with the user's short-term context data to obtain the user's fused memory context. The fused memory context is serialized and arranged into instructions to obtain the user's generated instruction sequence; The user's personalized response content is obtained by inferring the intent from the generated instruction sequence.

[0065] By deeply analyzing and optimizing the information structure and core content of memory fragments, we focus on the key information with the highest dynamic weight and the closest relationship with the user's current interaction needs. We extract the core concepts, key data, core demands, and other core elements contained in these information one by one, while eliminating redundant and secondary auxiliary information. We then systematically organize the extracted core elements to form focused memory elements that can accurately reflect the user's core memory needs, ensuring that each focused memory element directly points to the user's core needs and key memory information.

[0066] Using key memory elements as the core of retrieval, the system traverses all knowledge entries and memory information stored in the updated long-term memory, filters out relevant content that highly matches the key memory elements in terms of semantics, logic, and correlation, extracts structured text descriptions of these relevant contents, and organizes the extracted text information in an orderly manner according to the hierarchical structure of "core information - related information - supplementary information" to ensure that the text description fully covers the key memory content related to the key memory elements, while maintaining the coherence and logicality of the expression, ultimately forming the user's long-term memory text.

[0067] Collect short-term contextual data of the user's current interaction, including the dialogue content of the new round of interaction, interaction scenario information, and the user's recent operation behavior records. Perform correlation analysis on long-term memory text and this short-term contextual data to explore the semantic associations, logical connections and scenario adaptation points between the two. Based on these correlation points, construct a unified semantic field so that the historical memory information in the long-term memory text and the real-time interaction information in the short-term contextual data can be integrated and complemented to form a fused memory context that can comprehensively reflect the user's current interaction background and memory needs. This ensures that the contextual information includes both long-term accumulated memory content and fits the current interaction scenario.

[0068] All information in the fused memory context is logically sorted and hierarchically divided, clarifying the sequence, causal relationship, and primary and secondary relationship between information. According to the preset instruction generation specifications, the information in the fused memory context is transformed into a series of instructions with clear execution intentions and logical coherence. Each instruction corresponds to a specific response generation task. These instructions are serialized and arranged in the execution order to ensure that the instruction sequence can clearly and accurately guide the subsequent response content generation process, forming a rigorous and directly executable generation instruction sequence.

[0069] By deeply analyzing the core intent of each instruction in the generated instruction sequence and combining the logical connections between instructions, the system deduces the user's real needs and potential expectations in the current interaction scenario. Based on the deduction results, the system calls the response generation module of the cognitive AI hardware to transform the memory information and intent corresponding to the instruction sequence into natural, fluent language content that fits the user's expression habits. This ensures that the response content not only accurately addresses the user's core needs but also incorporates personalized information from long-term memory, while conforming to the contextual logic of the current interaction. Ultimately, this results in personalized response content that meets the user's individual needs.

[0070] The beneficial effects are that by accurately extracting key memory elements from optimized memory fragments, combining them with updated long-term memory and short-term contextual data to construct a fused memory context, and then generating personalized responses through instruction serialization and intent deduction, a deep integration of long-term memory information and real-time interaction needs is achieved. This effectively improves the personalization, accuracy, and relevance of the response content, and solves the problem of low relevance between response content and user's long-term memory and current needs in traditional methods, providing users with a more targeted and practical interactive experience.

[0071] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0072] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for constructing and generating long-term memory for cognitive AI hardware, characterized in that, The method includes: S1. Perform semantic understanding on the user's interaction information to obtain the user's key fact data, and perform multi-dimensional analysis on the key fact data to obtain the user's multi-dimensional memory features; S2. Adaptively match and deduce the multidimensional memory features with the preset long-term memory knowledge graph to obtain the user's fusion strategy, and based on the fusion strategy, integrate the user's memory subgraph into the long-term memory knowledge graph to obtain the user's updated long-term memory. S3. Perform a two-way embedded retrieval on the key fact data and the updated long-term memory to obtain the bidirectional mapping relationship of the user; S4. Based on the bidirectional mapping relationship, perform long-term memory similarity retrieval on the user's new round of interaction information to obtain the user's associated memory fragments; S5. Dynamically assign weights to the associated memory fragments to obtain the user's optimized memory fragments, including: The associated memory fragments are deconstructed using multidimensional features to obtain the feature spectrum of the associated memory fragments. The weighted influence factors are extracted from the intrinsic correlation dimension of the feature spectrum to obtain the initial influence factors of the associated memory fragments; The initial influence factor is subjected to multi-source information collaborative deduction to obtain the enhanced influence factor of the associated memory fragment; The enhanced influence factors are subjected to contextual consistency fusion to obtain the dynamic weights of the associated memory segments; Based on the dynamic weights, the information structure of the associated memory fragments is reconstructed to obtain the user's optimized memory fragments; S6. Based on the optimized memory fragment, the updated long-term memory is adapted and deduced to obtain the user's personalized response content.

2. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, The process involves semantic understanding of the user's interaction information to obtain the user's key factual data, and then performing multi-dimensional analysis on this key factual data to obtain the user's multi-dimensional memory features, including: The system acquires user interaction information, which includes voice stream text, dialogue sequences, image information content, and emotional state parameters. Context-aware semantic deconstruction is performed on the interaction information to obtain the user's basic fact elements; The basic fact elements are categorized and merged to obtain the user's key fact data; Perform cross-dimensional feature mapping on the key fact data to obtain the user's primary characteristics; The primary features are fused to obtain the user's multidimensional memory features.

3. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, The process of adaptively matching and deriving the multidimensional memory features with a preset long-term memory knowledge graph to obtain the user's fusion strategy, and integrating the user's memory subgraph into the long-term memory knowledge graph based on the fusion strategy to obtain the user's updated long-term memory, includes: The multidimensional memory features are used to construct a feature graph to obtain the user's memory subgraph; Structural topology analysis is performed on the memory subgraph and the long-term memory knowledge graph to obtain the structural matching degree of the user. Based on the structural matching degree, attribute compatibility identification is performed on the memory subgraph to obtain the attribute association degree of the user; A multi-dimensional decision derivation is performed on the structural matching degree and the attribute correlation degree to obtain the user's fusion strategy; Based on the fusion strategy, the memory subgraph is reconstructed in relational topology to obtain the user's enhanced memory subgraph; The enhanced memory subgraph and the long-term memory knowledge graph are integrated and optimized to obtain the user's updated long-term memory.

4. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 3, characterized in that, The process of performing multi-dimensional decision derivation on the structural matching degree and the attribute correlation degree to obtain the user's fusion strategy includes: By performing state correlation between the structural matching degree and the attribute correlation degree, the user's decision characteristics are obtained; Based on the decision characteristics, heuristic deduction is performed on the strategy generation of the memory subgraph to obtain the user's strategy derivation path; The strategy derivation path is backtracked and verified to obtain the user's optimized strategy; The optimized strategy is instantiated and adapted to obtain the user's fusion strategy.

5. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, The step of performing a two-way embedded retrieval of the key fact data and the updated long-term memory to obtain the bidirectional mapping relationship of the user includes: The key fact data is feature-encoded to obtain the memory feature identifier of the key fact data; Based on the memory feature identifier, the entries in the updated long-term memory are matched and identified to obtain the user's bound memory entries; The memory feature identifier is matched with the bound memory entry to obtain the index association representation of the user; Based on the index association representation, bidirectional entry writing is performed on the user's index storage area to obtain the user's bidirectional mapping relationship.

6. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, Based on the bidirectional mapping relationship, the long-term memory similarity retrieval of the user's new round of interaction information is performed to obtain the user's associated memory fragments, including: Perform interactive semantic analysis on the user's new round of interaction information to obtain the user's query semantic features; Based on the bidirectional mapping relationship, the query semantic features are transformed into a vector space to obtain the user's query vector; Based on the query vector, a proximity search and matching is performed on the updated long-term memory to obtain the user's associated memory fragments.

7. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, The formula for calculating the dynamic weight is as follows: ； in, For the dynamic weight, For the first The aforementioned enhancing influence factors, For the first The contextual consistency coefficient of each of the associated memory segments, For the first The feature dimension weights of the associated memory segments The mean of the enhanced influence factor, The mean of the contextual consistency coefficients of the associated memory segments. This is a global correction factor. The current interaction time. For the first The generation time of the aforementioned associated memory fragments, The time decay coefficient, This represents the total number of the associated memory fragments.

8. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, The step of extracting weighted influence factors from the intrinsic correlation dimension of the feature spectrum to obtain the initial influence factors of the associated memory fragment includes: The associated memory fragments are metrically analyzed to obtain the activity and significance metrics of the associated memory fragments; Cross-validation is performed on the activity measure and the significance measure to obtain the validated association factor of the associated memory fragment; The importance of the verified association factors is ranked to obtain the ranking factors of the associated memory segments; The sorted factors are structured and encapsulated to obtain the initial influence factors of the associated memory fragments.

9. The method for constructing and generating long-term memory for cognitive AI hardware as described in claim 1, characterized in that, The process of adapting and extrapolating the updated long-term memory based on the optimized memory fragment to obtain the user's personalized response content includes: The optimized memory fragments are refined to obtain the user's focus memory elements; Based on the aforementioned focal memory elements, a structured text description is extracted from the updated long-term memory to obtain the user's long-term memory text. A semantic field is constructed by combining the long-term memory text with the user's short-term context data to obtain the user's fused memory context. The fused memory context is serialized and arranged into instructions to obtain the user's generated instruction sequence; The user's personalized response content is obtained by inferring the intent from the generated instruction sequence.

Images