Hierarchical memory management method, session interaction data query method, system and device

Abstract

The embodiment of the invention relates to the technical field of data processing, and discloses a hierarchical memory management method, a session interaction data query method, system and device and a storage medium, and the method comprises the steps: obtaining vehicle-mounted data information and session data information; based on the vehicle-mounted data information and the session data information, forming short-term memory information from the data information acquired in the first preset time; analyzing the short-term memory information, and forming middle-term memory information through clustering; and based on the medium-term memory information, judging whether the medium-term memory information meets a preset storage condition, and if yes, forming and storing long-term memory information. According to the technical scheme, the problems that in the prior art, due to the fact that conversation interaction lacks an effective multi-round conversation association mechanism, conversation context information cannot be reserved, memory limitation is obvious in the face of cross-period long-term conversations, and historical requirements are frequently forgotten are solved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing technology, specifically to a hierarchical memory management method, a session interaction data query method, a system, a device, and a storage medium. Background Technology

[0002] With the rapid development of automotive intelligence, cockpit functions have expanded from "convenient vehicle control" to diverse scenarios such as entertainment and lifestyle services. Users' core needs for the cockpit have also evolved to "intelligent driving and riding experience tailored to personalized needs and usage scenarios." However, significant technical bottlenecks still exist in the current application of large-scale models and intelligent agents within the cockpit: Regarding dialogue memory management, existing systems employ a single-turn, independent task processing model, lacking an effective multi-turn dialogue association mechanism. This results in the inability to retain dialogue context information, exhibiting significant memory limitations when facing long-term dialogues spanning different time periods, frequently leading to the forgetting of historical needs and severely disrupting dialogue continuity. In terms of scenario integration, existing technologies fail to effectively integrate in-vehicle data, unable to adjust services based on real-time usage scenarios or deeply mine user behavior data, resulting in highly homogenized service recommendations and insufficient personalized experiences. These issues collectively constrain the improvement of the interactive experience and service quality optimization of intelligent cockpit systems. Summary of the Invention

[0003] In view of the above problems, embodiments of the present invention provide a hierarchical memory management method, a conversation interaction data query method, a system, a device, and a storage medium to solve the problem that the conversation interaction in the prior art lacks an effective multi-turn dialogue association mechanism, which leads to the inability to retain dialogue context information, obvious memory limitations when facing long-term dialogues across time periods, and frequent forgetting of historical needs.

[0004] According to one aspect of the present invention, a hierarchical memory management method is provided, the method comprising: Acquire vehicle data and session data; Based on the vehicle data information and the session data information, the data information acquired within a first preset time period will be used to form short-term memory information; Analyze the short-term memory information and form medium-term memory information through clustering; Based on the intermediate memory information, it is determined whether it meets the preset storage conditions. If it does, it is formed into long-term memory information and stored.

[0005] In one alternative approach, forming short-term memory information from the data acquired within a first preset time period includes the following sub-steps: Store the most recent rounds of dialogue in the form of a triplet of question, answer, and timestamp; A sliding window mechanism is used to set the threshold for the number of conversations; When the number of conversations exceeds the conversation number threshold, the earliest conversation will be transferred to the intermediate memory information.

[0006] In one alternative approach, analyzing the short-term memory information and forming medium-term memory information through clustering includes the following sub-steps: Analyze the dialogue from which the short-term memory information is transferred; Determine the semantic continuity between the dialogue and the historical dialogue and update the dialogue list accordingly; Based on the updated dialogue list, generate summary information; The summary information is vectorized, and the session segments are formed based on vector semantic matching and clustering as medium-term memory information for storage.

[0007] In one alternative approach, determining the semantic continuity between the dialogue and historical dialogues and updating the dialogue list includes the following sub-steps: Determine the semantic continuity between the dialogue and historical dialogues; If continuity exists, the new conversation is added to the conversation list of that conversation; If no continuity exists, create a new dialog list and store it.

[0008] In one alternative approach, determining whether the intermediate-term memory information meets preset storage conditions, and if so, forming and storing long-term memory information, includes the following sub-steps: Based on the mid-term memory information, the cumulative number of visits, total number of conversations, and recent usage of the conversation segment are counted. The popularity score is calculated based on the cumulative number of visits, total number of conversations, and recent usage. Determine whether the heat value exceeds a preset heat threshold; If the preset popularity threshold is exceeded, the session will be input into the large model and a new user profile will be generated or an existing user profile will be updated. The user profile will be stored as long-term memory information.

[0009] In one alternative approach, inputting the session into a large model and generating a new user profile or updating an existing user profile includes the following sub-steps: Determine if a current user profile already exists; If the current user profile exists, the current user profile and the session segment are input into the large model for analysis and the user profile is updated. If the current user profile does not exist, the session segment is input into the large model for analysis and a new user profile is created.

[0010] According to a second aspect of the present invention, a method for querying session interaction data is provided, the method comprising: In response to the user's interactive command, the system queries the short-term memory information, medium-term memory information, and long-term memory information to obtain the target query information. Acquire in-vehicle short-term storage data and in-vehicle long-term storage data; The vehicle's short-term storage data, vehicle's long-term storage data, and target query information are integrated to form a comprehensive information package, which is then input into the large model. The large model performs inference based on the comprehensive information package, generates response results, and feeds them back to the user.

[0011] In one alternative approach, the in-vehicle short-term storage data includes vehicle status data and recent user operation data; The long-term in-vehicle storage data includes vehicle usage behavior patterns and scenario preferences.

[0012] In one alternative approach, acquiring the vehicle's short-term storage data and long-term storage data includes the following sub-steps: After acquiring user behavior data, system operating status data, and vehicle operating status data and performing preprocessing, the in-vehicle short-term storage data is formed. Based on the vehicle's short-term storage data, preset features are extracted to form the vehicle's long-term storage data.

[0013] According to a third aspect of the present invention, a hierarchical memory management system is provided, comprising: The communication module is used to acquire vehicle data information and session data information; The parsing module is used to form short-term memory information from the data information acquired within a first preset time period based on the vehicle data information and the session data information; and to analyze the short-term memory information to form medium-term memory information through clustering. The execution module is used to determine whether the intermediate memory information meets the preset storage conditions. If it does, it forms long-term memory information and stores it.

[0014] According to a fourth aspect of the present invention, a hierarchical memory management device is provided, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform operations as described above in the hierarchical memory management method and / or the session interaction data query method.

[0015] According to a fifth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the hierarchical memory management method and / or the session interaction data query method as described above.

[0016] (1) This invention constructs a hierarchical memory management system, which forms a progressive information processing mechanism through time-sensitive screening of short-term memory, semantic clustering of medium-term memory, and value assessment of long-term memory. This system integrates in-vehicle scenario data information and conversation data information to establish multi-dimensional memory associations, solving the problems of broken dialogue context and forgotten user needs in the prior art. At the same time, it realizes data support for personalized services through a dynamic memory accumulation mechanism.

[0017] (2) This invention achieves deep integration of conversation memory and vehicle data. By synchronously querying short-term, medium-term, and long-term memory information, it ensures the complete retrieval of dialogue context and key historical information, avoiding fragmented information loss caused by memory stratification. Combining short-term and long-term vehicle storage data provides a global input for the large model that covers both real-time environment and long-term features. Ultimately, the large model can generate response results that integrate user historical needs, real-time vehicle status, and long-term behavioral preferences based on multi-dimensional data cross-analysis, improving the adaptability and personalization of service recommendations and vehicle usage scenarios.

[0018] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a first embodiment of the hierarchical memory management method provided by the present invention is shown. Figure 2 The flowchart of a sub-step in step 120 of the first embodiment of the hierarchical memory management method provided by the present invention is shown. Figure 3 The flowchart of a sub-step in step 130 of the first embodiment of the hierarchical memory management method provided by the present invention is shown. Figure 4 The flowchart of a sub-step in step 140 of the first embodiment of the hierarchical memory management method provided by the present invention is shown. Figure 5 A flowchart illustrating a second embodiment of the session interaction data query method provided by the present invention is shown. Figure 6 A schematic diagram of an embodiment of the hierarchical memory management system provided by the present invention is shown; Figure 7 The flowchart illustrating the implementation of the hierarchical memory management method provided by this invention is shown. Detailed Implementation

[0020] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0021] Example 1: Figure 1 A flowchart of a first embodiment of the hierarchical memory management method of the present invention is shown, which is executed by the hierarchical memory management system. Figure 1 As shown, the method includes the following steps: Step 110: Obtain vehicle data information and session data information.

[0022] Among them, vehicle data information refers to vehicle operating status data and user operation behavior data, such as vehicle speed, gear position, air conditioning settings, and navigation operation records collected through vehicle sensors or cabin system logs. This feature is used to capture real-time vehicle usage scenario characteristics and provide vehicle status context for memory management.

[0023] The conversation data information refers to the dialogue text and related information generated by the user's interaction with the in-vehicle system, such as the text transcribed from the user's voice commands and the interaction time recorded by the speech recognition module. This feature is used to extract the semantic content of the dialogue and support multi-turn dialogue correlation analysis.

[0024] Specifically, it acquires vehicle data and session data to achieve multi-dimensional data fusion, providing a multi-source input foundation for subsequent memory management.

[0025] Step 120: Based on vehicle data information and session data information, form short-term memory information from the data information acquired within the first preset time period.

[0026] The first preset time refers to a time window parameter used to define the validity period of short-term memory storage. Specifically, the time range can be dynamically adjusted using a sliding window mechanism, for example, set to data within the most recent 30 minutes. This feature is used to filter outdated information, retain valid dialogue fragments, and prevent the accumulation of invalid data.

[0027] Specifically, based on vehicle data and conversation data, the data acquired within a first preset time period is used to form short-term memory information, and valid dialogue fragments are retained through timeliness filtering.

[0028] Step 130: Analyze short-term memory information and form medium-term memory information through clustering.

[0029] Clustering refers to a technique for semantically associating short-term memory information. Specifically, it can be achieved using a vector space model combined with similarity threshold matching, such as using the BERT model to generate text vectors and then calculating cosine similarity. This feature is used to integrate fragmented dialogue segments into conversational units with thematic continuity.

[0030] Specifically, short-term memory information is analyzed, and medium-term memory information is formed through clustering, thus solving the problem of integrating fragmented information.

[0031] Step 140: Based on the intermediate memory information, determine whether it meets the preset storage conditions. If it does, form long-term memory information and store it.

[0032] The preset storage conditions refer to the evaluation criteria for determining whether information has long-term preservation value. Specifically, this can be achieved by setting access frequency thresholds and semantic integrity indicators, such as requiring a session segment to be accessed more than 5 times and contain a complete expression of the request. This feature is used to filter high-value information and prevent invalid data from occupying storage resources for extended periods.

[0033] Specifically, based on intermediate memory information, it is determined whether it meets the preset storage conditions. If it does, it is formed into long-term memory information and stored, thereby realizing the persistent storage of core knowledge.

[0034] Understandably, the setting of the first preset time ensures the timeliness of short-term memory, clustering enhances the correlation of information, and the storage condition judgment mechanism establishes the standard for knowledge accumulation. The three work together to form a progressive memory optimization system.

[0035] The above methods construct a hierarchical memory management system, forming a progressive information processing mechanism through short-term memory time-sensitivity screening, medium-term memory semantic clustering, and long-term memory value assessment. This system integrates in-vehicle scenario data and conversation data to establish multi-dimensional memory associations, solving the problems of broken dialogue context and forgotten user needs in existing technologies. Simultaneously, it provides data support for personalized services through a dynamic memory accumulation mechanism.

[0036] Please combine Figure 2 , Figure 2 A flowchart illustrating a sub-step in step 120 of the hierarchical memory management method of the present invention is shown.

[0037] In step 120, the data information acquired within the first preset time period is used to form short-term memory information, including the following sub-steps: Step 121: Store the most recent rounds of dialogue as a triple of question, answer and timestamp.

[0038] For example, each round of dialogue can be stored in the format {"Question":"User-inputted question", "Answer":"System-generated answer", "Timestamp":"2023-05-01 14:30:25"}.

[0039] Step 122: Use a sliding window mechanism to set the threshold for the number of dialogues.

[0040] For example, the dialogue threshold can be set to 100 rounds. When a new dialogue comes in, the sliding window will dynamically adjust to keep the latest 100 rounds of dialogue in short-term memory.

[0041] Step 123: When the number of dialogues exceeds the dialogue threshold, the earliest dialogue is transferred to the intermediate memory information.

[0042] For example, when the 101st round of dialogue begins, the system will automatically transfer the earliest 1st round of dialogue from short-term memory to medium-term memory, so that short-term memory always contains only the latest 100th round of dialogue.

[0043] In this embodiment, by storing dialogue information in a structured manner, the system can quickly retrieve and analyze dialogue content. The sliding window mechanism ensures that short-term memory always contains up-to-date and manageable dialogue information, avoiding excessive storage space consumption. Simultaneously, dialogue exceeding a threshold is transferred to medium-term memory, ensuring that important information is not lost and providing a data foundation for subsequent in-depth analysis and long-term memory formation. This method effectively balances storage efficiency and information integrity, improving the efficiency and accuracy of the system's dialogue memory management.

[0044] Please combine Figure 3 , Figure 3 A flowchart illustrating a sub-step in step 130 of the hierarchical memory management method of the present invention is shown.

[0045] In step 130, short-term memory information is analyzed, and medium-term memory information is formed through clustering, including the following sub-steps: Step 131: Analyze the dialogue that transfers information from short-term memory.

[0046] Among them, the analysis of the incoming dialogue can use natural language processing techniques to extract the dialogue topic and key entities, such as establishing a dialogue feature set through named entity recognition and keyword extraction.

[0047] Step 132: Determine the semantic continuity between the dialogue and the historical dialogue and update the dialogue list.

[0048] When determining semantic continuity, the semantic similarity between the new dialogue and the historical dialogue can be calculated. When the similarity exceeds a set threshold, it is determined to be a continuous dialogue. For example, a cosine similarity algorithm is used to match dialogue vectors, and the similarity threshold is set to 0.85. During the dialogue list update process, continuous dialogues are integrated into the same list node, while non-continuous dialogues generate independent list branches, thus forming a tree-like topology.

[0049] Step 133: Generate summary information based on the updated dialogue list.

[0050] When generating summary information, a text summarization model can be used to extract the core content of the dialogue, such as generating condensed text through a sequence-to-sequence model based on the Transformer architecture.

[0051] Step 134: Vectorize the summary information and form session segments based on vector semantic matching clustering as medium-term memory information storage.

[0052] In the vectorization process, the summarized text is input into a pre-trained language model to generate high-dimensional semantic vectors, such as using the BERT model to output 768-dimensional vectors. During clustering, hierarchical clustering algorithms are used to aggregate semantically similar vectors into conversation segments, for example, setting a clustering distance threshold of 1.2 to achieve topic consistency.

[0053] In this embodiment, when a new dialogue enters, its core content is first extracted through semantic analysis, and its similarity is calculated with the node at the end of the current dialogue list. If the similarity reaches a threshold, the dialogue is appended to the end of the existing list; otherwise, a new branch of the list is created. After each list update, a summary is generated for all dialogue content in the current list, extracting condensed text containing user intent and key information. The summary text is then converted into semantic vectors, and distance calculations within the vector space are used to group summary vectors with similar semantics into the same conversation segment. For example, in a car navigation scenario, multiple consecutive rounds of dialogue about "finding a charging station" are clustered into the same conversation segment, while a subsequent dialogue that suddenly switches to "playing music" forms a new conversation segment.

[0054] In some embodiments, determining the semantic continuity between a dialogue and historical dialogues and updating the dialogue list includes the following sub-steps: Determine the semantic continuity between the dialogue and the historical dialogue; If continuity exists, the new conversation is added to the conversation list of that conversation; If no continuity exists, create a new dialog list and store it.

[0055] Semantic continuity is determined by calculating the similarity between the current dialogue vector and the latest node in the historical dialogue list. For example, a cosine similarity threshold of 0.7 can be used; if the similarity exceeds the threshold, continuity is considered to exist. When updating the dialogue list, a time window mechanism is used to limit the list length. For example, only dialogue nodes associated within the last 5 minutes are retained, and nodes outside this time range are automatically removed. The creation of a new list can be determined by combining semantic similarity and time interval. For example, independent storage is triggered when the similarity is below the threshold and the interval between adjacent dialogues exceeds 2 minutes.

[0056] Specifically, when a new dialogue is transferred from short-term memory, its vector representation is extracted using a semantic model and compared with the latest node in all current dialogue lists for similarity. If the node similarity of a certain list reaches a preset threshold, the dialogue is appended to the end of that list, and the list's timestamp is updated; if none of the lists meet the condition, an independent list carrying a new timestamp is generated. By dynamically maintaining a set of semantically consistent lists, dialogues on the same topic are identified as consecutive conversation segments during the clustering phase, while unrelated dialogues are isolated to different lists. Therefore, conversation segments in medium-term memory can accurately reflect the real dialogue topics when generating summaries and vectorization, avoiding information mixing caused by incorrect clustering, and providing a reliable data foundation for calculating the popularity value of long-term memory and updating user profiles.

[0057] Please combine Figure 4 , Figure 4 A flowchart illustrating a sub-step in step 140 of the hierarchical memory management method of the present invention is shown.

[0058] In step 140, based on the intermediate memory information, it is determined whether it meets the preset storage conditions. If it does, long-term memory information is formed and stored, including the following sub-steps: Step 141: Based on the mid-term memory information, count the cumulative number of visits, total number of conversations, and recent usage of the conversation segment.

[0059] The cumulative number of visits is defined as the number of times a user actively accesses the session segment, such as through voice commands or touch operations; the total number of dialogues includes the total number of rounds of user questions and system responses within the session segment, for example, a single dialogue involving five rounds of questions and answers is counted as five; the degree of recent use is calculated using a time decay function, for example, using an exponential decay model to give higher weight to access behavior within the last three days.

[0060] Step 142: Calculate the popularity score based on the cumulative number of visits, total number of conversations, and recent usage.

[0061] The popularity value can be calculated using a weighted summation method, such as 40% for the number of visits, 30% for the total number of conversations, and 30% for the recent usage. When the weighted sum exceeds a set threshold, subsequent processing is triggered.

[0062] Step 143: Determine whether the heat value exceeds the preset heat threshold.

[0063] The preset heat threshold is dynamically adjusted according to the vehicle usage scenario. For example, the threshold is set to 70 in commuting mode and 60 in long-distance driving mode.

[0064] Step 144: If the preset popularity threshold is exceeded, the session is input into the large model and a new user profile is generated or an existing user profile is updated.

[0065] When a threshold is triggered, the conversation segment is input into a large model, where user behavior features are extracted through semantic parsing. For example, the pattern of "navigating to the gym after get off work every Friday" is extracted, and a user profile containing usage time preferences and destination distribution is further generated.

[0066] Step 145: Store the user profile as long-term memory information.

[0067] In this embodiment, when a user engages in multi-turn dialogue with the in-vehicle system, the system first clusters conversation segments in the medium-term memory information. The cumulative number of visits is updated in real time; for example, a music recommendation conversation segment may be visited three times. The total number of dialogues counts the number of interaction rounds within the conversation segment, for example, including eight rounds of question and answer. The recent usage level is calculated based on the decay value of the last visit time. If the weighted calculation result of the three factors reaches a preset threshold, for example, a weighted value of 75, the large model analysis process is triggered. After the conversation segment is input into the large model, user preference keywords are identified through natural language processing, such as "classical music" and "nighttime playback," thereby updating the music type preference field in the user profile. The updated user profile is stored as long-term memory; for example, "music preference: classical music, playback time: 20:00-22:00" is written to the database.

[0068] Therefore, in subsequent interactions, the system can directly access long-term memory information to achieve personalized service recommendations, such as automatically playing classical music playlists at night. By dynamically quantifying multi-dimensional indicators and correlating them with large-scale model analysis, high-value conversation content is effectively filtered, ensuring that long-term memory information accurately reflects user needs and improving the timeliness and adaptability of service recommendations.

[0069] In some embodiments, the session is input into a large model and a new user profile is generated or an existing user profile is updated, including the following sub-steps: Determine if a current user profile already exists; If a current user profile exists, the current user profile and the session segment are input into the large model for analysis and the user profile is updated. If the current user profile does not exist, the session segment will be input into the large model for analysis and a new user profile will be created.

[0070] Specifically, a hash table can be used to index the mapping relationship between user IDs and profile storage paths. When a current user profile exists, historical profile data and the current session segment can be input into the large model via concatenation or separate channel input. When a current user profile does not exist, the triggering condition for creating a new user profile can be further verified by combining the session segment's popularity value. During the user profile update process, the large model can use an incremental learning mechanism to semantically supplement historical profiles.

[0071] In this embodiment, when determining whether to generate or update a user profile based on the popularity value of mid-term memory information, the system first checks the user identifier to determine if there is an associated profile storage record. If a record exists, historical profile data and the current session segment are input into the large model. For example, the JSON format data of the historical profile is concatenated with the current session text, enabling the large model to recognize the continuous changes in user preference features. If no record exists, an initial profile is generated directly based on the current session segment. During this process, the popularity value calculation module and the session segment clustering module work together. For example, the user profile generation or update operation is triggered only when the session segment clustering result meets the preset semantic density and the popularity value exceeds a threshold, avoiding redundant processing of low-value session data.

[0072] Example 2: Figure 5 A flowchart of a second embodiment of the session interaction data query method of the present invention is shown. Figure 5 As shown, the method includes the following steps: Step 210: In response to the user's interactive command, query short-term memory information, medium-term memory information and long-term memory information to obtain the target query information.

[0073] The short-term memory information stores the most recent rounds of dialogue through a sliding window mechanism. When the number of dialogues exceeds a threshold, the earliest dialogue is transferred to the medium-term memory information. The medium-term memory information forms conversation segments through semantic clustering. The popularity value of the long-term memory information is calculated based on the cumulative number of visits, the total number of dialogues, and the recent use time decay factor.

[0074] Specifically, when a user issues a voice command, the most recent five rounds of dialogue are prioritized for retrieval. Simultaneously, the semantic matching module searches for relevant conversation segments in the mid-term memory, such as recalling dialogue clusters related to "air conditioning adjustment." User profile data, including historically set temperature preference ranges, is extracted synchronously from the long-term memory.

[0075] Step 220: Obtain vehicle short-term storage data and vehicle long-term storage data.

[0076] The short-term in-vehicle storage data includes real-time vehicle speed, air conditioning set temperature, and the user's touch operation records in the last 10 minutes; the long-term in-vehicle storage data includes the user's monthly commuting route selection frequency and seasonal seat heating preference modes.

[0077] Specifically, the vehicle's current gear position, remaining driving range, and steering wheel heating status are collected in real time via the CAN bus and correlated with the user's seat ventilation usage records during the afternoon hours over the past 30 days.

[0078] Step 230: Integrate short-term vehicle storage data, long-term vehicle storage data, and target query information to form a comprehensive information package and input it into the large model.

[0079] Specifically, all data is input into a large model after feature alignment and dimensional unification. The model fuses multi-source information through a cross-attention mechanism, such as comparing the real-time outside temperature with the user's long-term preferred temperature difference threshold, and combining the remaining mileage of the current navigation destination to generate suggested air conditioning settings and seat ventilation intensity.

[0080] Step 240: The large model performs inference based on the comprehensive information package, generates response results, and feeds them back to the user.

[0081] The comprehensive information package is encapsulated in JSON format and includes a timestamp, data source identifier, and normalized feature vector.

[0082] Specifically, while the response result is broadcast through the TTS module, the key parameters of this interaction are updated to the short-term memory window, the relevant semantic features are supplemented to the medium-term memory cluster center, and the temperature preference parameters in the user profile are recalculated based on the latest operation.

[0083] The above methods achieve deep integration of conversation memory and in-vehicle data. By synchronously querying short-term, medium-term, and long-term memory information, complete retrieval of dialogue context and key historical information is ensured, avoiding fragmented information loss caused by memory hierarchies. Combining short-term and long-term in-vehicle storage data provides a global input for the large model, covering both real-time environment and long-term features. Ultimately, the large model can generate response results that integrate user historical needs, real-time vehicle status, and long-term behavioral preferences based on multi-dimensional data cross-analysis, improving the adaptability and personalization of service recommendations to different vehicle usage scenarios.

[0084] In some embodiments, short-term in-vehicle storage data includes vehicle status data and recent user operation data; long-term in-vehicle storage data includes vehicle usage behavior patterns and scenario preferences.

[0085] In this embodiment, vehicle status data and recent user operation data are aligned with timestamps to form a dynamic data stream. Vehicle usage behavior patterns and scenario preferences are used to generate a static feature library through feature extraction algorithms. The dynamic data stream and the static feature library are input into a large model through vector concatenation.

[0086] In some embodiments, acquiring in-vehicle short-term storage data and in-vehicle long-term storage data includes the following sub-steps: After acquiring and preprocessing user behavior data, system operation status data, and vehicle operation status data, in-vehicle short-term storage data is formed.

[0087] Based on the short-term vehicle storage data, preset features are extracted to form long-term vehicle storage data.

[0088] User behavior data is collected via onboard sensors, including user operation frequency, touch trajectory, and voice command duration. System operating status data is obtained via the CAN bus, acquiring in-vehicle entertainment system operating parameters and energy consumption data. Vehicle operating status data is collected via inertial measurement unit, including acceleration, steering angular velocity, and braking pressure values. Preprocessing includes data cleaning, format conversion, and timestamp alignment. Pre-defined feature extraction employs a sliding time window mechanism; for example, it counts the number of high-frequency user operations at 10-minute intervals, calculates the average acceleration change rate at 50-kilometer intervals, and analyzes the distribution of air conditioning usage periods at 24-hour intervals.

[0089] In this embodiment, the touch trajectory in the user behavior data is preprocessed to generate a two-dimensional coordinate sequence. A dynamic time warping algorithm is used to match a preset operation template to identify high-frequency operation behaviors. The braking pressure value in the vehicle's operating status data is calculated through piecewise integration to generate a braking habit feature vector. The preprocessed short-term storage data undergoes dimensionality compression using a feature extraction engine. Long-term vehicle storage data is persistently stored using a feature vector database, with each feature associated with a timestamp and vehicle mileage information.

[0090] Example 3: like Figure 6 As shown, Figure 6 A schematic diagram of an embodiment of the hierarchical memory management system 60 of the present invention is shown. The hierarchical memory management system 60 includes: a communication module 61, a parsing module 62, and an execution module 63.

[0091] The communication module 61 is used to acquire vehicle data information and session data information. The parsing module 62 is used to form short-term memory information based on the vehicle data information and session data information acquired within a first preset time period; analyze the short-term memory information and form medium-term memory information through clustering. The execution module 63 is used to determine whether the medium-term memory information meets the preset storage conditions; if it does, it forms long-term memory information and stores it.

[0092] Specifically, after vehicle data is input through communication module 61, the data is temporarily stored in the cache queue of parsing module 62. Parsing module 62 extracts data in 5-minute time windows, generates a short-term memory information matrix containing keyword weights after normalization, and then integrates the discrete data into medium-term memory information blocks with common semantic features through clustering algorithms. Execution module 63 evaluates each information block in real time. If its access frequency reaches 3 times within 10 seconds and its correlation degree exceeds 0.75, a compression encoding operation is triggered to reduce the dimensionality of the data and store it in non-volatile memory.

[0093] The above system constructs a hierarchical memory management system, forming a progressive information processing mechanism through short-term memory time-sensitivity filtering, medium-term memory semantic clustering, and long-term memory value assessment. This system integrates in-vehicle scenario data and conversation data to establish multi-dimensional memory associations, solving the problems of fragmented dialogue context and forgotten user needs in existing technologies. Simultaneously, it provides data support for personalized services through a dynamic memory accumulation mechanism.

[0094] Example 4: Figure 7 The diagram shows a structural schematic of an embodiment of the hierarchical memory management device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the hierarchical memory management device.

[0095] like Figure 7 As shown, a hierarchical memory management device may include: a processor 701, a communication interface 702, a memory 703, and a communication bus 704.

[0096] The processor 701, communication interface 702, and memory 703 communicate with each other via communication bus 704. Communication interface 702 is used to communicate with other devices or network elements such as servers. The processor 701 executes program 710, which implements the steps in the hierarchical memory management method described above.

[0097] Specifically, program 710 may include program 710 code, which includes computer-executable instructions.

[0098] Specifically, the processor 701 may be a central processing unit 701 CPU, an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The hierarchical memory management device includes one or more processors 701, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0099] Memory 703 is used to store program 710. Memory 703 may include high-speed RAM memory 703, and may also include non-volatile memory 703, such as at least one disk storage 703.

[0100] Specifically, program 710 can be called by processor 701 to cause the hierarchical memory management device to perform the following operations: Acquire vehicle data and session data; Based on vehicle data and session data, the data acquired within the first preset time period will be used to form short-term memory information. Analyze short-term memory information and form medium-term memory information through clustering; Based on the intermediate memory information, it is determined whether it meets the preset storage conditions. If it does, it is formed into long-term memory information and stored.

[0101] In an alternative implementation, program 710 is invoked by processor 701 to cause the hierarchical memory management device to execute specific sub-steps of steps 120, 130, and 140 in embodiment 1.

[0102] The above devices construct a hierarchical memory management system, forming a progressive information processing mechanism through short-term memory time-sensitive filtering, medium-term memory semantic clustering, and long-term memory value assessment. This system integrates in-vehicle scenario data and conversation data to establish multi-dimensional memory associations, solving the problems of broken dialogue context and forgotten user needs in existing technologies. At the same time, it provides data support for personalized services through a dynamic memory accumulation mechanism.

[0103] Example 5: This invention provides a computer-readable storage medium storing at least one executable instruction. When the executable instruction is executed on a hierarchical memory management device, the hierarchical memory management device performs the hierarchical memory management method and / or the session interaction data query method in any of the above method embodiments.

[0104] Specifically, the executable instructions can be used to cause the hierarchical memory management device to perform the following operations: Acquire vehicle data and session data; Based on vehicle data and session data, the data acquired within the first preset time period will be used to form short-term memory information. Analyze short-term memory information and form medium-term memory information through clustering; Based on the intermediate memory information, it is determined whether it meets the preset storage conditions. If it does, it is formed into long-term memory information and stored.

[0105] By constructing a hierarchical memory management system, a progressive information processing mechanism is formed through short-term memory time-sensitive screening, medium-term memory semantic clustering, and long-term memory value assessment. This system integrates in-vehicle scenario data and conversation data to establish multi-dimensional memory associations, solving the problems of broken dialogue context and forgotten user needs in existing technologies. At the same time, it provides data support for personalized services through a dynamic memory accumulation mechanism.

[0106] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Furthermore, the embodiments of this invention are not directed to any particular programming language.

[0107] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. Similarly, for the sake of brevity and to aid in understanding one or more aspects of the invention, in the description of exemplary embodiments of the invention above, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of the invention.

[0108] Those skilled in the art will understand that the modules in the device of the embodiment can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiment can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

[0109] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A hierarchical memory management method characterized by, The method comprises the following steps: acquiring vehicle data information and session data information; forming short-term memory information from data information acquired within a first preset time based on the vehicle data information and the session data information; analyzing the short-term memory information to form medium-term memory information through clustering; judging whether the medium-term memory information meets a preset storage condition, and forming long-term memory information and storing it if it meets the condition.

2. The hierarchical memory management method of claim 1, wherein, The step of forming short-term memory information from data information acquired within a first preset time comprises the following sub-steps: storing the latest several rounds of conversations in the form of triples of questions, answers and timestamps; setting a conversation threshold value by using a sliding window mechanism; when the number of conversations exceeds the conversation threshold value, transferring the earliest conversation to the medium-term memory information.

3. The hierarchical memory management method of claim 1, wherein, The step of analyzing the short-term memory information to form medium-term memory information through clustering comprises the following sub-steps: analyzing the conversation transferred by the short-term memory information; judging the semantic continuity of the conversation with historical conversations and updating a conversation link list; generating summary information based on the updated conversation link list; vectorizing the summary information, clustering the conversation segments based on vector semantic matching, and storing the conversation segments as medium-term memory information.

4. The hierarchical memory management method of claim 3, wherein, The step of judging the semantic continuity of the conversation with historical conversations and updating a conversation link list comprises the following sub-steps: judging the semantic continuity of the conversation with historical conversations; if there is continuity, adding the new conversation to the conversation link list of the conversation; if there is no continuity, creating a new conversation link list and storing it.

5. The hierarchical memory management method of claim 1, wherein, The step of judging whether the medium-term memory information meets a preset storage condition, and forming long-term memory information and storing it if it meets the condition comprises the following sub-steps: based on the conversation segments of the medium-term memory information, counting the cumulative access times, the total number of conversations and the recent usage degree of the conversation segments; based on the cumulative access times, the total number of conversations and the recent usage degree, calculating a heat value; judging whether the heat value exceeds a preset heat threshold value; if the heat value exceeds the preset heat threshold value, inputting the conversation into a large model and generating a new user portrait or updating an existing user portrait; storing the user portrait as long-term memory information.

6. The hierarchical memory management method of claim 5, wherein, The step of inputting the conversation into a large model and generating a new user portrait or updating an existing user portrait comprises the following sub-steps: judging whether there is a current user portrait; if there is the current user portrait, inputting the current user portrait and the conversation segment into the large model for analysis and completing user portrait updating; if there is no current user portrait, inputting the conversation segment into the large model for analysis and completing new user portrait creation.

7. A method of conversational interaction data query, the method comprising: The method comprises the following steps: in response to an interaction instruction issued by a user, querying the short-term memory information, the medium-term memory information and the long-term memory information to obtain target query information; acquiring vehicle short-term storage data and vehicle long-term storage data; integrating the vehicle short-term storage data, the vehicle long-term storage data and the target query information to form a comprehensive information package and inputting the comprehensive information package into the large model; the large model performs reasoning based on the comprehensive information package to generate a response result and feeds back the response result to the user.

8. The conversational interaction data query method of claim 7, wherein, The vehicle short-term storage data comprises vehicle state data and user recent operation data; The vehicle-mounted long-term storage data includes vehicle behavior patterns and scene preferences.

9. The conversational interaction data query method of claim 7, wherein, The obtaining of the vehicle-mounted short-term storage data and the vehicle-mounted long-term storage data includes the following sub-steps: After obtaining user behavior data, system running state data and vehicle body running state data and preprocessing, the vehicle-mounted short-term storage data is formed; Based on the vehicle-mounted short-term storage data, after extracting preset features, the vehicle-mounted long-term storage data is formed.

10. A hierarchical memory management system characterized by, It comprises: A communication module for obtaining vehicle-mounted data information and session data information; An analysis module for forming short-term memory information based on the vehicle-mounted data information and the session data information, and the data information obtained within a first preset time; Analyzing the short-term memory information and forming medium-term memory information through clustering; An execution module for judging whether the medium-term memory information meets the preset storage condition based on the medium-term memory information, and forming long-term memory information and storing if it meets the condition.

11. A hierarchical memory management device, characterized by, It comprises: A processor, a memory, a communication interface and a communication bus, which complete mutual communication through the communication bus; The memory is used to store at least one executable instruction, and the executable instruction makes the processor execute to realize the operations in the hierarchical memory management method and / or the session interaction data query method in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, wherein the device where the computer readable storage medium is located executes the computer program to realize the hierarchical memory management method and / or the session interaction data query method in any one of claims 1 to 9.

Images