Data processing method and electronic equipment

Abstract

The invention provides a data processing method and electronic equipment. The method comprises the following steps: in response to a target triggering condition, obtaining first data corresponding to the target triggering condition; the first data represents interaction data between the user and the intelligent agent; the target triggering condition represents a condition that the user indicates that the first data is stored as memory data; generating target memory of the user according to the first data; the target memory represents the first type of memory; the first type of memory characterizes a memory generated based on a user active trigger.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method and an electronic device. Background Technology

[0002] With the development of Artificial Intelligence (AI) technology, Large Language Models (LLMs) have been widely used in various application scenarios. However, agent systems based on LLMs face the following problems when interacting with users over long periods: First, the problem of memory loss. Because traditional agent systems lack persistent memory mechanisms, they cannot retain users' historical information and preferences between multiple conversations, making each interaction feel like the first meeting, severely impacting the user experience. Second, the difficulty of maintaining memory consistency. When user information changes, existing agent systems struggle to effectively update and maintain the consistency of user-related memories, leading to memory conflicts (i.e., information conflicts). Summary of the Invention

[0003] In view of the above problems, this application provides a data processing method and an electronic device for intelligent registration and management of user memories.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a data processing method, the method comprising: In response to a target triggering condition, first data corresponding to the target triggering condition is acquired; the first data represents the interaction data between the user and the intelligent agent; the target triggering condition represents the condition under which the user instructs the first data to be stored as memory data. Based on the first data, the user's target memory is generated; the target memory represents the first type of memory; the first type of memory represents the memory generated based on the user's active triggering.

[0005] Additionally, this application also provides an electronic device, comprising: a memory and a processor; wherein the memory stores a computer program, and when the processor executes the computer program, it is used for: In response to a target triggering condition, first data corresponding to the target triggering condition is acquired; the first data represents the interaction data between the user and the intelligent agent; the target triggering condition represents the condition under which the user instructs the first data to be stored as memory data. Based on the first data, the user's target memory is generated; the target memory represents the first type of memory; the first type of memory represents the memory generated based on the user's active triggering.

[0006] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this application. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0008] Figure 1 A schematic diagram illustrating the implementation flow of a data processing method provided in this application; Figure 2 A schematic diagram illustrating an exemplary process for obtaining first data provided in this application; Figure 3 A schematic diagram illustrating an exemplary process for generating a target memory is provided for this application; Figure 4 A schematic diagram illustrating an exemplary implementation process for updating newly added memories to already stored memories, provided for the purposes of this application; Figure 5 This application provides an exemplary flowchart for updating stored memories in a target database. Figure 6 A line-of-sight flowchart of an embodiment provided in this application; Figure 7 A schematic diagram illustrating the implementation process of another embodiment provided in this application; Figure 8 A schematic diagram illustrating the implementation process of another embodiment provided in this application; Figure 9 This is a schematic diagram of the hardware entity of an electronic device provided in this application. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. The terms "first / second / third" are merely used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first / second / third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terminology used herein is for the purpose of describing this application only and is not intended to limit this application.

[0010] To address the issues of memory loss and difficulty in maintaining memory consistency in intelligent agent systems, the following solutions have been proposed: Solution 1 provides users with basic memory storage functionality. This solution only performs simple storage of user memories and lacks complex memory management mechanisms. Solution 2 relies on context windows during interaction to maintain short-term memories relevant to the user. This solution cannot maintain long-term memories relevant to the user, thus still affecting the user's interaction experience.

[0011] Based on this, this application provides a data processing method that can be executed by an electronic device. The electronic device can be a laptop, tablet, desktop computer, set-top box, mobile device (e.g., mobile phone, portable music player, personal digital assistant, dedicated messaging device, portable gaming device), or a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing 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.

[0012] The technical solutions in this application will now be clearly and completely described with reference to the accompanying drawings.

[0013] Figure 1 This application provides a schematic diagram of the implementation flow of a data processing method, as shown below. Figure 1 As shown, the method includes the following steps S11 to S12: Step S11: In response to the target triggering condition, acquire the first data corresponding to the target triggering condition; the first data represents the interaction data between the user and the intelligent agent; the target triggering condition represents the condition under which the user instructs the first data to be stored as memory data.

[0014] Here, an intelligent agent can refer to an artificial intelligence entity capable of perceiving its environment, making autonomous decisions, and executing actions to complete a specified task. An intelligent agent can be of any type. In some embodiments, the intelligent agent can be a software intelligent agent, such as a chatbot, recommendation system, or automation program. In some embodiments, the intelligent agent can be a physical robot, such as an industrial robot or service robot. In some embodiments, the intelligent agent can be an intelligent agent used to perform any function, such as an intelligent agent used for information processing and generation, an intelligent agent used for strategy planning, or an intelligent agent used for interaction and collaboration with humans or other intelligent agent systems. In some embodiments, the intelligent agent can be an intelligent agent used to generate memory data.

[0015] The first data refers to any type of interaction data determined during the interaction between the user and the intelligent agent. In some implementations, the first data may be user input data, such as text data, image data, audio data, video data, multimodal data, Uniform Resource Locator (URL), or user operation flow for a specified task. In some implementations, the first data may be data generated by the intelligent agent, such as task processing results (e.g., text generation results, image generation results, video generation results, etc.) generated or organized by the intelligent agent based on user input data.

[0016] The target trigger condition refers to the condition in which the user instructs the first data to be stored as memory data; that is, the condition initiated by the user to store the first data as user memory.

[0017] In some implementations, the target triggering condition can be any condition in which the user instructs the first data to be stored as memory data. In some implementations, the target triggering condition can be that during interaction with the intelligent agent, the user specifies that the first data should be stored as memory data via preset text keywords, gestures, voice commands, etc. For example, the user performs a preset memory storage gesture to store the first data corresponding to that gesture as memory data. Another example is that the user inputs preset text keywords (e.g., "remember this text") to store the first data corresponding to that text keyword as memory data. In some implementations, the target triggering condition can be that during interaction with the intelligent agent, the user triggers an operation involving a specified function key or combination of function keys; wherein the specified function key can be a physical function key or a virtual function key in the user interface (UI) (i.e., a specified control in the UI); the specified function key combination can be a combination of multiple physical function keys, or a combination of physical and virtual function keys, etc.

[0018] In some implementations, after determining the target triggering condition, the first data corresponding to the target triggering condition can be determined and obtained in any way.

[0019] In some implementations, when the target trigger condition is text or voice content input by the user, the first data corresponding to the target trigger condition can be determined based on the semantics of the text or voice; alternatively, keyword detection can be performed on the text or voice to determine the first data corresponding to the target trigger condition based on the keywords. For example, when the target trigger condition is the user-input text "Remember the interaction content in the previous hour", the interaction content in the most recent hour can be used as the first data based on the time information in the target trigger condition.

[0020] In some implementations, the first data corresponding to the target triggering condition can be determined based on the triggering form of the target triggering condition. For example, if the target triggering condition is a preset memory start gesture and a memory end gesture, the interaction content between the memory start gesture and the memory end gesture can be determined as the first data. As another example, if the target triggering condition is operation information for a preset function key, the operation time for the preset function key can be used as the start time for receiving the first data.

[0021] Step S12: Generate the user's target memory based on the first data; the target memory represents the first type of memory; the first type of memory represents the memory generated based on the user's active triggering.

[0022] Here, after obtaining the first data corresponding to the target triggering condition, data processing is performed on the first data to generate the user's target memory.

[0023] In some implementations, the user's target memory can be generated in any way.

[0024] In some implementations, the first data and its corresponding metadata can be used as the user's target memory.

[0025] In some implementations, keyword extraction can be performed on the first data to obtain at least one entity keyword and / or at least one relational keyword contained in the first data, and then the obtained keywords can be stored as the user's target memory. For example, if the first data is the text "The steak at the Western restaurant is delicious" entered by the user, the entity keywords "Western restaurant" and "steak" and the relational keywords "made" and "delicious" can be extracted from it, and then the two entity keywords and two relational keywords can be stored in a relational database or a graph database to form the user's target memory.

[0026] In some implementations, the first data can be digested, and the resulting digest data and its corresponding metadata can be used as the user's target memory. For example, if the first data is an article that the user has read, a summary of the article can be generated, and the summary and its corresponding metadata can be used as the user's target data.

[0027] Type I memory refers to memory generated based on user-initiated triggers, i.e., memory generated according to user commands. Since Type I memory is generated based on explicit user actions, it can be called explicit memory. In some implementations, because Type I memory is actively created by the user, it possesses higher accuracy, value, and stability. When subsequently updating, retrieving, and / or maintaining the user's memory data, Type I memory can be assigned higher priority and updated using strategies distinct from other types of memory to improve its priority and stability within the user's memory data.

[0028] Since the target triggering condition corresponding to the target memory is the condition for the user to instruct the first data to be stored as memory data, the target memory is a memory generated based on the user's active triggering, that is, the target memory is a first type of memory.

[0029] In the data processing method provided in this application, firstly, in response to a target triggering condition, first data corresponding to the target triggering condition is acquired. The first data represents interaction data between the user and the intelligent agent, and the target triggering condition represents the condition under which the user instructs the first data to be stored as memory data. Then, based on the first data, a target memory for the user is generated. This target memory represents a first type of memory, which in turn represents a memory generated based on a user's active triggering. In this way, on the one hand, generating the user's memory data based on the interaction data between the user and the intelligent agent (e.g., the first data) allows the intelligent agent to better understand the background and context of the current interaction and the user's intent, thereby providing the user with an efficient and personalized interactive experience. On the other hand, compared to storing user memories using basic memory storage functions, this application can generate corresponding user memories not only by detecting the interaction content between the user and the intelligent agent but also by generating user memories based on the user's active triggering, thus providing a diversified and intelligent memory creation strategy.

[0030] In some implementations, the target triggering condition includes the detection of target semantic information.

[0031] Here, target semantic information represents the semantic information instructing the user to store the first data as memory data. For example, the target semantic information could be phrases like "remember this data," "remember," or "generate memory" from the user's input text or voice information.

[0032] In some implementations, semantic analysis can be performed on the user-inputted data to determine whether it contains the target semantics. For example, the input data can be converted into feature vectors, and then a classifier can be used for semantic analysis. Alternatively, deep learning-based algorithms can be used to perform semantic analysis on the input data, such as word vector representations, convolutional neural networks, pre-trained large language models, graph neural networks, and multimodal models.

[0033] Thus, as Figure 2 As shown, in some embodiments, step S11 above, namely, in response to the target triggering condition, obtaining the first data corresponding to the target triggering condition, can be implemented as the following steps S21 to S22: Step S21: Obtain the second data; the second data represents the interaction data generated during the interaction between the user and the intelligent agent.

[0034] Here, the second data refers to the interaction data generated during the interaction between the user and any type of intelligent agent. In some implementations, the second data can be user input data, such as text data, image data, audio data, video data, multimodal data, URLs, or the user's operation flow for a specified task. In some implementations, the second data can be data generated by the intelligent agent, such as the task processing results generated or organized by the intelligent agent based on the user's input data.

[0035] Step S22: In response to the second data containing target semantics, the target data corresponding to the target semantics in the second data is determined as the first data.

[0036] Here, semantic analysis is performed on the second data to determine whether it contains the target semantics. If the semantic analysis result of the second data contains the target semantics, it is considered that the target semantics has been detected in the second data, and the information of the detected target semantics is used as the target trigger condition, thereby determining the target data corresponding to the target semantics in the second data as the first data.

[0037] In some implementations, the target data corresponding to the target semantics can be a complete sentence containing the target semantics. For example, if the second data consists of multiple complete interactive statements input by the user, and the target semantics are detected in one of the complete sentences, then that complete sentence is used as the target data, such as "Remember I love eating apples".

[0038] In some implementations, the target data corresponding to the target semantics can be data determined based on a phrase containing the target semantics. For example, if the phrase containing the target semantics is "remember the above dialogue content", then all the interactive content generated in the current dialogue can be used as the target data; or, if the phrase containing the target semantics is "remember the following dialogue content", then all the interactive content that appears in the current dialogue after the target semantics can be used as the target data.

[0039] In the embodiments provided in this application, the target semantics in the second data are used to determine the first data, thereby enabling the generation of a first type of target memory based on the first data. This allows users to actively trigger the generation of the first type of memory using natural language, making the generation method of the first type of memory more in line with user interaction habits and possessing high flexibility and intelligence, thus achieving an efficient first type of memory (i.e., explicit memory) creation process.

[0040] In some implementations, the target triggering condition includes a user triggering a target control in the target user interface; the target user interface represents the user interaction interface provided by the target intelligent agent.

[0041] A target intelligent agent refers to an intelligent agent with memory generation capabilities. In some implementations, the target intelligent agent can be a general-purpose intelligent agent with memory generation capabilities, or it can be an intelligent agent specifically designed for memory generation.

[0042] In some implementations, the user interface provided by the target intelligent agent can be a dockable and hideable panel. For example, the user can move the display position of the target user interface on the screen by dragging, and the target user interface can automatically shrink to the edge of the screen and be hidden.

[0043] A target control is an interactive element used to actively trigger a memory creation operation during user interaction with a target intelligent agent. In some implementations, the target control can be placed anywhere on the target user interface, such as the upper half, lower half, or sidebar of the interface. In some implementations, the target control can be designed as any graphical button, such as a button with the text "Remember" or a button with a brain-shaped icon.

[0044] Here, the target triggering condition includes user triggering of the target control. Thus, when a user trigger action on the target control is detected, the data corresponding to that trigger action is used as the first data, and then the target memory corresponding to the first data is generated.

[0045] Thus, in some embodiments, step S11 above, namely, obtaining the first data corresponding to the target triggering condition in response to the target triggering condition, can be implemented as the following steps: In response to a user triggering a target control, the data corresponding to the triggering action is determined as the first data.

[0046] Here, when the user triggers the target control, the data corresponding to the triggering action is determined, and that data is then identified as the first data.

[0047] In some implementations, the data corresponding to the triggering action may refer to text, images, voice, or video data entered by the user in the target user interface after the user triggers the target control.

[0048] In some implementations, the data corresponding to the triggering action may refer to any text, image, voice, or video data selected in the display interface on the screen before or after the user triggers the target control.

[0049] In the embodiments provided in this application, users can create a first type of memory by triggering a target control, making the creation of the first type of memory more intuitive and convenient, thereby improving the user's memory creation experience.

[0050] like Figure 3 As shown, in some embodiments, step S12 above, namely, generating the user's target memory based on the first data, may include the following steps S31 to S32: Step S31: Perform digest processing on the first data to obtain the memory digest corresponding to the first data.

[0051] Performing summary processing on the first data refers to extracting core information, important conclusions, and / or key patterns from the first data to generate concise and easy-to-understand summary information, i.e., a memory summary corresponding to the first data. In some implementations, the memory summary corresponding to the first data can be generated in any manner.

[0052] In some implementations, summarization can be performed on the first data using an extractive approach. For example, a graph model can be used to perform summarization on the first data, where sentences in the first data are treated as nodes, and the similarity between sentences is used as the weights of the edges between nodes. A graph ranking algorithm is then used to calculate the importance of each node, and the summaries are composed of sentences with higher importance. Alternatively, a machine learning classification model can be used to perform summarization on the first data, where a classification model is used to determine whether each sentence in the first data should be included in the summary.

[0053] In some implementations, generative methods can be used to perform summarization processing on the first data. For example, a deep learning model can be used to model the first data, learn its semantic information, and thus generate a memory summary corresponding to the first data. Another example is using a large language model to perform semantic analysis on the first data, thereby generating a memory summary corresponding to the first data.

[0054] Step S32: Determine the memory summary corresponding to the first data as the target memory and add a target tag to the target memory; wherein, the target tag is used to indicate that the target memory belongs to the first type of memory.

[0055] After determining the memory summary corresponding to the first data, this memory summary is identified as the target memory corresponding to the first data. Simultaneously, a target tag is added to this target memory to indicate that it belongs to the first type of memory. Therefore, based on this target tag, the memory type of the target memory can be quickly determined. In some implementations, this target tag can be "user-triggered memory," "user-actively recorded," or "explicit memory," etc.

[0056] In the embodiments provided in this application, by performing summary processing on the first data, on the one hand, the amount of user memory data can be effectively reduced and the retrieval efficiency of memory data can be improved; on the other hand, invalid or duplicate information can be filtered out, avoiding the storage of a large amount of redundant content. In addition, by adding a target tag to the target memory, the target memory can be identified as a memory generated based on user-initiated triggering, which has a higher priority and stability. In this way, during subsequent updates, retrieval, and / or maintenance of user memory data, the memory type and update strategy of the target memory can be quickly identified based on the target tag, thereby achieving more refined memory management.

[0057] In some embodiments, the first data includes a target image and target text; the memory summary corresponding to the first data includes a first summary and a second summary; Step S31 above, namely, performing digest processing on the first data to obtain the memory digest corresponding to the first data, may include the following steps S311 to S312: Step S311: Generate an image description corresponding to the target image based on the target text; Step S312: Generate a first summary and a second summary based on the image description and the target text, respectively; the first summary has a higher level of information detail than the second summary.

[0058] Here, when the first data includes image type data (i.e., target image) and text type data (i.e. target text), two memory summaries are generated for the first data, namely, a first summary and a second summary, and the information detail of the first summary is higher than that of the second summary.

[0059] Information detail refers to the amount of information related to the first data contained in the memory summary. In some embodiments, information detail refers to the number of keywords in the first data contained in the memory summary. In some embodiments, information detail refers to the semantic detail related to the first data contained in the memory summary.

[0060] Here, when generating the first summary and the second summary corresponding to the first data, firstly, image features are extracted from the target image to generate an image description corresponding to the target image; then, the first summary and the second summary are generated based on the image description and the target text.

[0061] Because the target image contains a wealth of information, such as the positional relationships between different objects, color features, and brightness characteristics, generating an image description based solely on the first image would result in a large amount of information in the generated description, including much irrelevant information that the user is not interested in. Furthermore, since both the target text and the target image are primary data, they are highly likely to have a semantic connection. Therefore, when generating the image description information for the target image, semantic analysis is performed on the target text, and image features are extracted from the target image based on the semantics of the target text to obtain the corresponding image description. This ensures that the image description contains the key information that the user wants to remember. For example, if the target image is a photo of a parked car taken by the user, and the target text is "Remember the parking space," this target text information indicates that the user is primarily focused on the parking space information when generating the corresponding image description.

[0062] After determining the image description corresponding to the target image, the image description and the target text are summarized to generate a first summary and a second summary. In some embodiments, the same or different summarization methods can be used to generate the first summary and the second summary.

[0063] In the embodiments provided in this application, when the first data includes both image-type data and text-type data, generating a first summary and a second summary with different levels of information detail can enrich the expression of user memory and adapt to different memory application scenarios. For example, the first summary with higher information detail can be used for subsequent interactive data generation, while the second summary with lower information detail can be used for subsequent memory retrieval and displaying key memories to the user. In this way, the integrity of the target memory is guaranteed while also taking into account memory retrieval efficiency and flexibility.

[0064] like Figure 4 As shown, in some embodiments, after performing steps S11 to S12 as described above, the data processing method further includes the following steps S41 to S42: Step S41: If the similarity between the user's newly added memory and the stored memory is higher than the first threshold, the newly added memory and the stored memory are recombined to obtain the updated memory.

[0065] Here, "stored memories" refers to historical memories generated based on the user's historical interaction records with the agent, and which have been stored in the user's memory database (e.g., the target database). "New memories" refers to memories generated after the aforementioned stored memories, based on the user's interaction records with the agent, and which have not yet been stored in the user's memory database (e.g., the target database).

[0066] Before storing a new memory in the user's target database, it is necessary to determine the storage method of the new memory based on its similarity to the user's historical memories, so as to avoid storing multiple memories with high similarity in the target database.

[0067] Thus, if the similarity between the newly added memory and any stored historical memory is not higher than the first threshold, the newly added memory is stored in the user's target database; conversely, if the similarity between the newly added memory and the stored memory is higher than the first threshold, the newly added memory and the stored memory are recombined to obtain an updated memory, and the updated memory is stored in the user's target database.

[0068] The condition that the similarity between a newly added memory and a stored memory is higher than a first threshold means that the semantic similarity between the two memories is higher than the first threshold. The first threshold is any pre-set similarity threshold, such as 0.7, 0.8, or 0.9. For example, if the newly added memory is "I like to eat apples" and the stored memory is "I like to eat fruit," it can be determined that both the newly added memory and the stored memory describe the user's dietary preferences and have a high degree of similarity.

[0069] In some implementations, stored memories with a similarity higher than a first threshold to the newly added memory can be determined using any method. In some implementations, a large language model can be used to perform semantic analysis on the newly added memory and each historical memory to determine stored memories with a similarity higher than the first threshold. In some implementations, first, the vector representations of the newly added memory and each historical memory are determined; then, the cosine similarity between the newly added memory and each historical memory is determined by calculating the vector distance (e.g., cosine distance); finally, the historical memory with a cosine similarity higher than the first threshold is taken as the stored memory corresponding to the newly added memory.

[0070] Reorganization refers to integrating newly added memories with existing memories to generate an updated memory that contains key information from both. This reorganization process not only preserves important content from both the newly added and existing memories but also enhances the completeness and accuracy of the memory through intelligent semantic fusion.

[0071] Step S42: If the newly added memory belongs to the first type of memory, then add a target marker for the updated memory.

[0072] When reorganizing newly added memories and stored memories, if the newly added memory belongs to the first type of memory, the memory update strategy corresponding to the first type of memory is used to reorganize the newly added memory and the stored memory. Specifically, a target marker is added to the updated memory corresponding to the newly added memory and the stored memory, so as to indicate that the updated memory is a memory generated based on user-initiated triggering.

[0073] In some implementations, if the newly added memory belongs to the first type of memory, then according to the memory update strategy corresponding to the first type of memory, the updated memory corresponding to the newly added memory and the stored memory includes the complete content corresponding to the newly added memory, or the semantics of the newly added memory are not weakened in the updated memory.

[0074] In the embodiments provided in this application, on the one hand, by reorganizing newly added memories and stored memories, redundant information in user memory data (e.g., the user's target database) can be reduced and the coherence of memories can be improved, thereby improving the accuracy and efficiency of user memory retrieval; on the other hand, when the newly added memory belongs to the first type of memory, a target tag is added to the updated memory according to the memory update strategy corresponding to the first type of memory, thereby marking the updated memory as the first type of memory. In this way, the updated memory can be managed according to the management strategy of the first type of memory in the subsequent memory management process, so that the newly added memory generated based on the user's active trigger still has a high priority and stability.

[0075] In some embodiments, after performing steps S11 and S13 as described above, the data processing method further includes the following steps: When a user's new memory is semantically related to an existing memory, the new memory and the existing memory are recombined to obtain an updated memory; where the existing memory belongs to the first type of memory, the updated memory includes semantic content related to the existing memory.

[0076] Stored memories refer to historical memories generated based on the user's historical interaction records with the agent, and which have been stored in the user's memory database (e.g., a target database). Stored memories belong to the first type of memory, meaning they are memories generated based on user-initiated triggers and possess high accuracy, value, and stability. New memories refer to memories generated after the aforementioned stored memories, based on the user's interaction records with the agent, and which have not yet been stored in the user's memory database.

[0077] In some implementations, the semantic relevance between the newly added memory and the stored memory can mean that the similarity between the newly added memory and the stored memory is higher than the first threshold mentioned above, or that the newly added memory and the stored memory have semantic conflicts. For example, if the newly added memory is "the user likes to eat apples" and the stored memory is "the user does not like to eat apples", it can be seen that the newly added memory and the stored memory conflict in describing the user's dietary preferences.

[0078] Here, when there is a stored memory in the user's historical memory that is semantically related to the new memory, the new memory and the stored memory are recombined to obtain the updated memory.

[0079] In some implementations, if the similarity between the newly added memory and the stored memory is higher than a first threshold, the newly added memory and the stored memory are recombined to obtain an updated memory.

[0080] Since the stored memory is of type I memory and has high accuracy, in some embodiments, when reorganizing the new memory and the stored memory, the updated memory includes semantic content related to the stored memory; or, the updated memory includes the complete semantic content of the stored memory. That is, the memory reorganization operation does not weaken or delete any memory content of the type I memory. Furthermore, when the new memory is of type I memory, the updated memory includes semantic content related to both the new memory and the stored memory; or, the updated memory includes the complete semantic content of both the new memory and the stored memory.

[0081] Since stored memories are classified as Type I memories, possessing high accuracy, value, and stability, if the semantics of a newly added memory conflict with those of a stored memory, the updated memory corresponding to both the newly added and stored memories will still include semantic content related to the stored memory. In some implementations, when recombining newly added and stored memories, the content of the stored memory is simplified, and the simplified stored memory is used as part of the updated memory. In some implementations, when recombining newly added and stored memories, the complete semantic content of the stored memory is retained in the updated memory. For example, if the newly added memory is "I like to eat apples" and the stored memory is "I don't like to eat apples," recombining the newly added and stored memories yields the updated memory "The user likes to eat apples, but didn't like to eat apples before."

[0082] In the embodiments provided in this application, when the stored memory belongs to the first type of memory, the memory is updated according to the memory update strategy of the first type of memory (that is, the semantic information related to the stored memory is retained in the updated memory corresponding to the newly added memory and the stored memory). This can avoid the crude deletion of the first type of memory, thereby enabling the user's memory change process to be traced and improving the stability of the first type of memory in the user's memory storage system.

[0083] In some implementations, after recombining the new memory with the stored memory to obtain the updated memory, a content integrity check is performed on the updated memory. This content integrity check refers to verifying whether the updated memory has semantic truncation. For example, if the updated memory is "user likes," it is considered to have a semantic incompleteness problem. If the updated memory passes the content integrity check, it is stored in the user's memory database (e.g., the target database); otherwise, a new memory is regenerated based on the interaction data corresponding to the new memory.

[0084] like Figure 5 As shown, in some embodiments, after performing steps S11 and S12 as described above, the data processing method further includes the following steps S51 to S54: Step S51: Determine the weight of the stored memory based on the type of the user's stored memory.

[0085] Memory types can be categorized according to the way memories are created or the content of those memories.

[0086] In some implementations, memories can be categorized into first-type memories and second-type memories based on their creation method. As mentioned above, first-type memories are generated based on user-initiated triggers; second-type memories are not generated based on user-initiated triggers, but rather obtained through active memory extraction of user-agent interaction data. Therefore, first-type memories are user-specified stored memories, while second-type memories are created by the agent or its corresponding model after semantic analysis of user-related interaction data. Consequently, first-type memories have higher accuracy, importance, and stability than second-type memories. Therefore, when assigning weights to stored memories, first-type memories are assigned higher weights than second-type memories.

[0087] In some implementations, memories can be categorized according to their content into personal information type memories, skill type memories, preference type memories, health status type memories, and event type memories. Personal information type memories refer to memories related to the user's basic information, such as the user's name, age, and contact information; skill type memories refer to memories related to the user's own skills, such as memories related to musical instrument playing skills; preference type memories refer to memories related to the user's preferences in specific matters, such as dietary preferences and clothing preferences; health status type memories refer to memories related to the user's health status, such as being healthy or ill; and event type memories refer to memories related to events experienced by the user.

[0088] It is evident that, for users, the stability and importance of personal information type memory, skill type memory, preference type memory, health status type memory, and event type memory vary. Therefore, corresponding weights can be assigned to these different types of memory according to their stability and importance. For example, the weights of personal information type memory, skill type memory, preference type memory, health type memory, and event type memory decrease in that order.

[0089] In some implementations, for each stored memory, a weight is determined when the stored memory is generated or created. In some implementations, when generating the stored memory using a large language model, the large language model can simultaneously generate the weight of the stored memory based on its creation method or content. In some implementations, the weight of the stored memory can be determined based on its memory type and a pre-established mapping relationship between memory types and weights.

[0090] Step S52: Determine the target score of the stored memory based on the weight of the stored memory, the frequency of access to the stored memory within a specified time period, and the last access time of the stored memory.

[0091] The access frequency of a stored memory within a specified time period refers to the frequency with which the stored memory is retrieved and used within a specified time period (e.g., one year, one month). A higher access frequency for a stored memory indicates a higher level of importance; conversely, a lower access frequency indicates a lower level of importance.

[0092] The last access time to a stored memory refers to the time when the stored memory was last retrieved and used. For example, the last access time to the stored memory might be one day ago or one month ago. Therefore, the longer the time since the last access to a stored memory is from the current moment, the less impact and importance the stored memory has on the user; conversely, the longer the time since the last access to the stored memory is from the current moment, the stronger the impact and the higher the importance of the stored memory.

[0093] Thus, based on the weight of the stored memory, the frequency of access to the stored memory within a specified time period, and the time of the last access to the stored memory, a quantitative value that can characterize the value of the stored memory, namely, the target score, can be calculated.

[0094] Based on the above descriptions of the weight of stored memories, the frequency of access to stored memories within a specified time period, and the last access time of stored memories, we can see that: the higher the weight of a stored memory, the higher the target score corresponding to that stored memory; the higher the frequency of access to a stored memory within a specified time period, the higher the target score corresponding to that stored memory; and the shorter the time between the last access time of a stored memory and the current time, the higher the target score corresponding to that stored memory.

[0095] In some implementations, the target score can be calculated using the following formula (1). R : (1); in: I , representing the weights that have been stored in memory; t This represents the time interval between the last access time to a stored memory and the current time. f This indicates the frequency of access to stored memories within a specified time period; λ t , which represents the time decay coefficient; the default value of this time decay coefficient is 0.03, and it can be reset according to user requirements; β This indicates the frequency of access. f The weighting factor for this access frequency; the default value of the weighting factor is 0.2, and it can be reset according to user requirements.

[0096] Step S53: In response to the target score being less than the second threshold, the stored memory is deleted from the target database.

[0097] Here, the second threshold is a score threshold used to determine whether a stored memory is considered a forgotten memory. The higher the second threshold, the faster the user's memory database (i.e., the target database) is updated; conversely, the lower the second threshold, the slower the target database is updated. In some implementations, this second threshold can be set by the user or it can be a system default value.

[0098] After determining the target score of the stored memory, if the target score is less than a preset second threshold, it indicates that the stored memory is of low importance, and therefore it is deleted from the target database as a forgotten memory.

[0099] In step S54, in response to the target score being not less than the second threshold and less than the third threshold, the stored memory is identified as the memory to be reorganized; the third threshold is greater than the second threshold.

[0100] The third threshold is a fractional threshold used to determine whether to include stored memories as memories to be reconstructed. A higher third threshold results in a higher frequency of reconstructed stored memories and a faster update speed for the target database; conversely, a lower third threshold results in a lower frequency of reconstructed stored memories and a slower update speed for the target database. In some implementations, this third threshold can be set by the user or it can be a system default value.

[0101] After determining the target score of the stored memory, if the target score is not less than the second threshold and less than the third threshold, it means that the stored memory does not need to be forgotten, but needs to be reorganized with other memories in order to reduce the amount of data in the target database and improve the speed of subsequent memory retrieval.

[0102] In the embodiments provided in this application, by deleting stored memories with target scores lower than the second threshold, memories of lower importance in the target database can be cleaned up, the amount of data in the target database can be reduced, and the memory retrieval speed can be improved. At the same time, by reorganizing stored memories with target scores not less than the second threshold but less than the third threshold (i.e., setting a reorganization state between deleting memories and retaining the original memories), the information gap caused by deleting memories all at once can be avoided, and a gradual memory update strategy can be realized.

[0103] In some embodiments, after performing steps S11 and S12 as described above, the data processing method further includes the following steps: The second type of memory is obtained from the interaction data between the user and the agent; the second type of memory represents a memory that is not generated based on the user's active triggering; the retention strength of the first type of memory is greater than the retention strength of the second type of memory; the retention strength is used to determine whether the memory can be reorganized or deleted.

[0104] Here, the second type of memory refers to memory not generated based on active user triggering. In some implementations, the second type of memory is generated after actively retrieving interaction data between the user and the agent. The user is unaware of this memory generation process; therefore, the second type of memory can also be called implicit memory. For example, for interaction data between the user and the agent over a specified period, a large language model is used to perform semantic analysis on the interaction data to determine whether it contains data of a specified type, and then a second type of memory is generated based on this specified type of data. The specified type of data may include personal information type data, skill type data, preference type data, health status type data, or event type data, etc.

[0105] Retention strength refers to the stability of a memory. Since Type I memory is generated based on user-initiated events, while Type II memory is obtained through automatic retrieval of user-agent interaction data, it's clear that Type I memory is more accurate, important, and stable than Type II memory. Therefore, the retention strength of Type I memory is greater than that of Type II memory.

[0106] The retention strength is used to determine whether the memory can be reorganized or deleted. Because the retention strength of the first type of memory is higher than that of the second type of memory, in some embodiments, the first type of memory cannot be reorganized and / or deleted, while the second type of memory can be reorganized and / or deleted; or, the first type of memory can be reorganized but cannot be deleted, while the second type of memory can be reorganized and deleted; and so on.

[0107] In the embodiments provided in this application, the stability of the first type of memory in the user memory database (e.g., the target database) is improved by retaining intensity information, so that the information actively remembered by the user can be retained for a long time, thereby improving the accuracy of the user memory data.

[0108] Below, in conjunction with Figure 6 This application will now describe one embodiment of the application. In this embodiment, the user actively creates a first type of memory by triggering a target control in the target user interface provided by the target intelligent agent. For example... Figure 6 As shown, this embodiment includes the following steps S601 to S6010: Step S601: In response to the user triggering the target control, receive the input data input by the user; then, execute step S602. Step S602: Determine whether the input data contains image data; if yes, proceed to step S603; if no, proceed to steps S604 and S605. Step S603: Generate an image description corresponding to the image data; then, execute steps S604 and S605. If the input data only includes image data, a visual language model or a multimodal model is used to perform feature analysis on the image data to generate an image description corresponding to the image data. If the input data includes both image data and text data, the text data and image data are combined, and a multimodal model is used to perform feature analysis on the image data to generate an image description corresponding to the image data.

[0109] When generating an image description, the image description is generated in a system language supported by the electronic device running the data processing method.

[0110] Step S604: Generate detailed memory; then, proceed to step S606. Memorize in detail, corresponding to the first summary mentioned above.

[0111] When the input data contains only text data, a detailed memory is generated based on the text data; when the input data contains only image data, a detailed memory is generated based on the image description corresponding to the image data; when the input data contains both text data and image data, a detailed memory is generated based on the image description corresponding to the text data and the image data.

[0112] When generating detailed memory, the detailed memory is generated in a system language supported by the electronic device running the data processing method.

[0113] Step S605: Generate a concise memory; then, proceed to step S606. Concise memorization, corresponding to the second summary mentioned above.

[0114] When the input data contains only text data, a concise memory is generated based on the text data; when the input data contains only image data, a concise memory is generated based on the image description corresponding to the image data; when the input data contains both text data and image data, a concise memory is generated based on the image description corresponding to both the text data and the image data.

[0115] When generating a concise memory, the concise memory is generated in a system language supported by the electronic device running the data processing method.

[0116] Step S606: Determine the memory tag; then, proceed to step S607. Memory tags refer to tags for memory types. In some implementations, memory tags may include personal information, skills, preferences, health status, or events.

[0117] In some implementations, detailed memories can be input into a large language model or a classifier model to generate memory labels using the large language model or to determine memory labels using the classifier model.

[0118] Step S607: Determine if an image description exists; if yes, proceed to step S608; if no, proceed to step S6010. Here, if the input data contains image data and an image description corresponding to the image data is generated using the above step S603, it is determined that an image description exists; otherwise, it is determined that an image description does not exist.

[0119] Step S608: Determine the image description label; then, proceed to step S609. In some implementations, image descriptions can be input into a large language model or classifier model to determine image description labels.

[0120] Step S609: Save the image description and image description tag as an image description memory; then, execute step S6010. Step S6010: Save detailed memory and concise memory.

[0121] Here, when an image description exists, the identification information of the image description memory is associated with and stored in relation to the detailed memory and the concise memory when saving the detailed memory and the concise memory.

[0122] Below, in conjunction with Figure 7 Another embodiment provided in this application will be described below. In this embodiment, a second type of memory is generated by actively analyzing the interaction data between the user and the intelligent agent, and if target semantics are detected during the generation of the second type of memory, a first type of memory is generated. This embodiment includes the following steps S701 to S708: Step S701: Determine the interaction data to be generated from the historical interaction data; then, execute step S702. In some implementations, the user's historical interaction data with the agent and a predefined classification template are input into a large language model so that the large language model can determine whether the historical interaction data contains interaction data to be generated based on the classification template.

[0123] In some implementations, the classification template includes multiple memory types and their corresponding judgment information; wherein, the multiple memory types may be personal information types, skill types, preference types, health status types, event types, etc.; the judgment information refers to the interactive data used to determine whether the interactive data can generate the corresponding type of memory. For example, the judgment information corresponding to the personal information type includes the definition of personal information and which types of information personal information includes.

[0124] In this way, when the large language model determines that the historical interaction data includes judgment information that matches a certain memory type in the classification template, the interaction data is identified as the interaction data to be generated as the memory.

[0125] In some implementations, after determining the interaction data to be generated as a memory, the large language model determines whether to generate a user memory based on the data volume (e.g., the number of tokens) of the interaction data. Here, when the data volume is less than a data volume threshold, the interaction data representing the memory to be generated may be incomplete, and therefore no user memory is generated for it. When the data volume is greater than the data volume threshold, the interaction data to be generated is segmented according to a specified data length (e.g., 300 tokens) to obtain multiple data slices of the interaction data to be generated, and the following process is performed for each data slice. Specifically, when segmenting the interaction data to be generated as a memory, it is necessary to ensure that the beginning of each data slice has semantic coverage with the preceding data slice, and the end of each data slice has semantic coverage with the following data slice.

[0126] Step S702: Perform summary processing on the interactive data of the memory to be generated to obtain a memory summary; then, execute steps S703 and S707 respectively. A memory summary is a portion of a user's memory extracted from the interactive data to be generated. In some implementations, for each data slice in the interactive data to be generated, a summary processing is performed using a large language model to obtain the corresponding memory summary.

[0127] Step S703: Extract the user profile information from the memory summary; then, proceed to step S704. User profile information is a portion of user memories extracted from the interaction data to be generated.

[0128] Step S704: Determine whether the interactive data of the memory to be generated contains the target semantics; if yes, proceed to step S705; if no, proceed to step S706. In some implementations, the target semantics can be "remember this," "remember the above," "remember the below," etc., which can represent any semantic information that the user instructs to generate memory data.

[0129] Step S705: Based on the memory update strategy corresponding to the first type of memory, the generated memory summary and user profile information, determine the memory to be updated; then, proceed to step S708. Here, the memory update strategies corresponding to the first type of memory include the memory reorganization strategies, memory update strategies, and / or memory deletion strategies corresponding to the first type of memory described in any one or more steps above, which will not be elaborated here.

[0130] Step S706: Based on the memory update strategy corresponding to the second type of memory, the generated memory summary and user profile information, determine the memory to be updated; then, proceed to step S708. Here, the memory update strategies corresponding to the second type of memory include the memory storage, reorganization, and / or deletion strategies corresponding to the second type of memory described in any one or more steps above, which will not be repeated here.

[0131] Step S707: Generate memory tags corresponding to the memory summary; then, proceed to step S708. Here, memory tags refer to classification tags for memory summaries, or memory type tags for second-type memories. In some implementations, memory tags may include personal information, skills, preferences, health status, events, etc.

[0132] Step S708: Store the determined memories to be updated and their corresponding memory tags in the relational database and the vector database, and store the generated user profile information in the relational database.

[0133] Here, both relational databases and vector data belong to the target database mentioned above.

[0134] In some implementations, the relational database used to store update memories and the relational database used to store user profile information may be the same or different relational databases.

[0135] In some implementations, when the currently generated memory is recombined with the stored memory to obtain an updated memory, a rule-based verification is performed on the updated memory before storing it in the user's target database. That is, it is determined whether the fields such as the memory summary and user profile information corresponding to the updated memory are complete.

[0136] In some implementations, if the currently generated memory is a first type of memory and is recombined with an existing memory to obtain an updated memory, before storing the updated memory in the user's target database, it is verified whether the updated memory has a target tag; and / or, it is verified whether the memory content of the currently generated memory has been deleted, rewritten, or weakened in the updated memory.

[0137] Below, in conjunction with Figure 8 An embodiment for generating a first type of memory and performing user tasks based on the generated memory will be described.

[0138] First, the intelligent agent 810 receives the first information 801 input by the user; wherein, the first information 801 is "Hello, please remember that I like decaffeinated coffee"; Then, agent 810 performs semantic analysis on the first information 801 and determines that the first information 801 contains the target semantics of the first type of memory actively triggered by the user, namely "remember". Agent 810 generates a memory summary 802 and user profile information 803 corresponding to the first information 801. The memory summary 802 can be "the user likes decaffeinated coffee". At the same time, a target label is generated for the memory summary to identify that the memory summary is the first type of memory generated based on the user's active trigger. The user profile information 803 can be "preference: decaffeinated coffee".

[0139] Then, agent 820 receives second information 804 input by the user; wherein agent 820 may be the same as or different from agent 810; the second information 804 is "What should I order for Alice to drink?"; Subsequently, agent 820 determines the user task as “What does Alice like to drink” based on the second information 804; based on the second information 804 and the determined user task, agent 820 retrieves the user’s memory data to obtain memory summary 802 and user profile information 803. Finally, based on the second information 804, the determined user task, the memory summary 802, and the user profile information 803, the intelligent agent generates interactive data 805 to be fed back to the user, namely, "Hello, Alice likes Earl Grey tea. According to your personal memory, you like decaffeinated coffee. Would you like to order yourself a decaffeinated almond milk coffee?"

[0140] As can be seen from the above embodiments, during the execution of user tasks, retrieving the user's memory data can provide the user with more personalized interactive feedback information, thereby improving the user's interactive experience.

[0141] As can be seen from the above embodiments, the data processing method provided in this application, on the one hand, realizes full lifecycle management of user memory creation, updating, decay, and forgetting; on the other hand, it distinguishes the triggering method of memory creation into user-initiated triggering and non-user-initiated triggering, thereby generating first-type memory and second-type memory, and assigns different updating, decay, and forgetting strategies to first-type memory and second-type memory, realizing multi-level and intelligent management of user memory; in addition, user memory is divided into memory summary and user profile, and stored using vector database and relational database. In this way, multiple retrieval methods can be used to retrieve user memory data. For example, the best matching 25 (BM25) and hybrid search methods can be combined to retrieve user memory data, thereby improving the accuracy and efficiency of memory retrieval.

[0142] Based on the above embodiments, this application also provides an electronic device. For example... Figure 9 As shown, the electronic device 900 includes a memory 910 and a processor 920; wherein, the memory 910 stores a computer program, and when the processor 920 executes the computer program, it is used for: In response to a target triggering condition, first data corresponding to the target triggering condition is acquired; the first data represents the interaction data between the user and the intelligent agent; the target triggering condition represents the condition under which the user instructs the first data to be stored as memory data. Based on the first data, the user's target memory is generated; the target memory represents the first type of memory; the first type of memory represents the memory generated based on the user's active triggering.

[0143] In some implementations, the target triggering condition includes detecting target semantic information; The processor 920 is configured to: acquire second data; the second data represents interaction data generated during the interaction between the user and the intelligent agent; and, in response to the second data containing target semantics, determine the target data corresponding to the target semantics in the second data as the first data.

[0144] In some implementations, the target triggering condition includes a user triggering a target control in the target user interface; the target user interface represents the user interaction interface provided by the target agent. The processor 920 is used to respond to a user triggering a target control and determine the data corresponding to the triggering action as the first data.

[0145] In some implementations, the processor 920 is configured to: perform digest processing on the first data to obtain a memory digest corresponding to the first data; determine the memory digest corresponding to the first data as a target memory; and add a target tag to the target memory; wherein the target tag is used to indicate that the target memory belongs to a first type of memory.

[0146] In some implementations, the first data includes a target image and target text; the memory summary corresponding to the first data includes a first summary and a second summary; the processor 920 is configured to: generate an image description corresponding to the target image based on the target text; generate a first summary and a second summary based on the image description and the target text, respectively; the information detail level of the first summary is higher than that of the second summary.

[0147] In some implementations, the processor 920 is further configured to: reconstruct the new memory and the stored memory to obtain an updated memory if the similarity between the new memory and the stored memory is higher than a first threshold; and add a target marker to the updated memory if the new memory belongs to a first type of memory.

[0148] In some implementations, the processor 920 is further configured to reorganize the newly added memory and the stored memory to obtain an updated memory when the newly added memory and the stored memory are semantically related; wherein, if the stored memory belongs to a first type of memory, the updated memory includes semantic content related to the stored memory.

[0149] In some implementations, the processor 920 is further configured to: determine the weight of a stored memory based on the type of the user's stored memory; determine a target score for the stored memory based on the weight of the stored memory, the frequency of access to the stored memory within a specified time period, and the last access time of the stored memory; delete the stored memory from the target database in response to the target score being less than a second threshold; and identify the stored memory as a memory to be reorganized in response to the target score being not less than the second threshold and less than a third threshold; wherein the third threshold is greater than the second threshold.

[0150] In some implementations, the processor 920 is further configured to obtain a second type of memory from user-agent interaction data; the second type of memory represents a memory not generated based on user-initiated triggering; the retention strength of the first type of memory is greater than the retention strength of the second type of memory; the retention strength is used to determine whether the memory can be reorganized or deleted.

[0151] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. In some embodiments, the functions or included units of the device provided in this application can be used to perform the methods described in the above method embodiments. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0152] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A data processing method, comprising: In response to a target triggering condition, first data corresponding to the target triggering condition is acquired; The first data represents the interaction data between the user and the intelligent agent; The target triggering condition represents the condition under which the user instructs the first data to be stored as memory data; Based on the first data, the user's target memory is generated; The target memory represents a first type of memory; the first type of memory represents a memory generated based on user-initiated triggers.

2. The method according to claim 1, wherein the target triggering condition includes detecting target semantic information; The step of obtaining first data corresponding to the target triggering condition in response to the target triggering condition includes: Obtain the second data; The second data represents the interaction data generated during the interaction between the user and the intelligent agent; In response to the second data containing the target semantics, the target data corresponding to the target semantics in the second data is determined as the first data.

3. The method according to claim 1, wherein the target triggering condition includes user triggering of a target control in the target user interface; the target user interface represents the user interaction interface provided by the target intelligent agent; The step of obtaining first data corresponding to the target triggering condition in response to the target triggering condition includes: In response to the user triggering the target control, the data corresponding to the triggering action is determined as the first data.

4. The method according to any one of claims 1 to 3, wherein generating the user's target memory based on the first data comprises: Perform digest processing on the first data to obtain the memory digest corresponding to the first data; The memory summary corresponding to the first data is identified as the target memory, and a target tag is added to the target memory; The target marker is used to indicate that the target memory belongs to the first type of memory.

5. The method according to claim 4, wherein the first data includes a target image and target text; the memory summary corresponding to the first data includes a first summary and a second summary; The step of performing digest processing on the first data to obtain a memory digest corresponding to the first data includes: Based on the target text, generate an image description corresponding to the target image; Based on the image description and the target text, a first summary and a second summary are generated respectively; the first summary has a higher level of information detail than the second summary.

6. The method according to claim 4, further comprising: If the similarity between the user's newly added memory and the stored memory is higher than a first threshold, the newly added memory and the stored memory are recombined to obtain an updated memory; If the newly added memory belongs to the first type of memory, then the target tag is added to the updated memory.

7. The method according to claim 1, further comprising: When the user's new memory is semantically related to the stored memory, the new memory and the stored memory are recombined to obtain an updated memory; wherein, if the stored memory belongs to the first type of memory, the updated memory includes semantic content related to the stored memory.

8. The method according to any one of claims 1 to 7, further comprising: The weight of the stored memory is determined based on the type of memory stored by the user. The target score of the stored memory is determined based on the weight of the stored memory, the frequency of access to the stored memory within a specified time period, and the time of the last access to the stored memory. In response to the target score being less than a second threshold, the stored memory is deleted from the target database; In response to the target score being not less than the second threshold and less than the third threshold, the stored memory is identified as a memory to be reconstructed; The third threshold is greater than the second threshold.

9. The method according to any one of claims 1 to 7, further comprising: Obtain second-type memories from user-agent interaction data; The second type of memory representation is not based on memories generated by user-initiated actions; The retention strength of the first type of memory is greater than that of the second type of memory; The retention strength is used to determine whether the memory can be reorganized or deleted.

10. An electronic device comprising a memory and a processor; wherein, The memory stores a computer program, and when the processor executes the computer program, it is used to: In response to a target triggering condition, first data corresponding to the target triggering condition is acquired; the first data represents interaction data between the user and the intelligent agent; the target triggering condition represents a condition in which the user instructs the first data to be stored as memory data. Based on the first data, the user's target memory is generated; The target memory represents a first type of memory; the first type of memory represents a memory generated based on user-initiated triggers.

Images