Method and system for realizing long-term memory of digital human based on mental model and large language model

Abstract

The invention discloses a method and a system for realizing long-term memory of a digital human based on a mental model and a large language model, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring sensing data through a sensor; preprocessing the sensing data to obtain preprocessed data; according to the preprocessed data, constructing short-term memory; acquiring long-term memory from the short-term memory; querying the short-term memory and the long-term memory according to the query content to obtain a query result; managing long-term memory according to a query result; the memory management efficiency can be improved, the memory is concluded, classified and sorted through the large language model, redundant information can be effectively reduced, and the memory management efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for realizing long-term memory of a digital human based on a mental model and a large language model. Background Art

[0002] With the rapid development of artificial intelligence technology, digital humans are increasingly being used in various fields. However, existing digital human systems have certain limitations in processing complex conversations and long-term memory. Traditional RAG creates retrieval by vectorizing external knowledge bases, creates associations through vector retrieval in conversations, and embeds relevant knowledge texts that should be associated into the context of large language models, so that relevant knowledge affects expression generation. However, this method is acceptable for knowledge question and answering, but it is difficult to let digital humans use knowledge in depth. It is difficult to meet users' needs for continuity and personalized services.

[0003] After building memory through mental models, we have a new option. We can first extract knowledge from the original knowledge and store the extracted knowledge in categories. Then, we can associate the extracted knowledge through the operation of different roles in the cognitive process. For example, the operation of the three cognitive roles of the curious, the answerer, and the inferer will be affected by causal knowledge, while the operation of the curious will be affected by the attention dimension.

[0004] In this way, compared with the traditional RAG model, knowledge is directly embedded in the dialogue to generate answers. In the mental model, knowledge first affects the cognitive process by affecting the cognitive-related roles in the cognitive model, changes the storage of cognitive conclusions in long-term storage, and then embeds the cognitive conclusion information into the dialogue to create associations. This is how humans use knowledge: using knowledge to influence cognitive thinking, thereby affecting dialogue. With the help of cognitive processes, we can go beyond the shallow use of knowledge in traditional RAG and achieve a deep application of knowledge like humans. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for realizing long-term memory of digital humans based on mental models and large language models. By combining mental models with the construction of multi-role cognitive models and the powerful natural language processing capabilities of large language models, long-term memory and in-depth application of information can be achieved, thereby improving the performance of digital humans in complex conversations and personalized services.

[0006] In order to solve the above technical problems, the present invention provides a method for realizing long-term memory of a digital human based on a mental model and a large language model, comprising the following steps:

[0007] Acquire sensory data through sensors;

[0008] Preprocessing the sensed data to obtain preprocessed data;

[0009] Construct short-term memory based on preprocessed data;

[0010] Access long-term memory from short-term memory;

[0011] Query the short-term memory and the long-term memory according to the query content to obtain the query result;

[0012] Manage long-term memory based on query results.

[0013] Preferably, the sensing data is preprocessed to obtain preprocessed data, which specifically includes the following steps:

[0014] Convert the sensory data into text description information and store the text description information as the sensory_data field;

[0015] The sensor_data fields of the sensor data collected at the same time are grouped together and timestamps are added as preprocessed data.

[0016] Preferably, constructing short-term memory based on the preprocessed data specifically includes the following steps:

[0017] The preprocessed data is merged to obtain short-term memory.

[0018] Preferably, acquiring long-term memory from short-term memory specifically comprises the following steps:

[0019] All short-term memories under session IDs that have not added new data for more than two days are used as long-term memories.

[0020] Preferably, querying the short-term memory and the long-term memory according to the query content to obtain the query result specifically includes the following steps:

[0021] Build your character;

[0022] Get the query content;

[0023] Associating and processing the query content according to the role to obtain a label;

[0024] According to the label, query is performed in short-term memory and long-term memory to obtain the query result.

[0025] Preferably, managing long-term memory according to the query result specifically includes the following steps:

[0026] When the tag query successfully matches the short-term memory, the corresponding short-term memory attention is increased;

[0027] When the label query and the on-stage memory are matched successfully, the corresponding long-term memory attention is increased;

[0028] At regular intervals, reduce the focus on short-term memory and long-term memory;

[0029] When the long-term memory attention is lower than the attention threshold, the corresponding long-term memory is deleted.

[0030] Preferably, each time a query is successfully matched, the corresponding short-term memory attention or long-term memory attention is increased by one point;

[0031] Every other day, the short-term memory focus or long-term memory focus will decrease by one point.

[0032] Preferably, the method further comprises the following steps:

[0033] According to the long-term memory attention, the corresponding long-term memory indicator weight is obtained;

[0034] According to the long-term memory and the indicator weights, the short-term memory is scored to obtain the initial short-term memory attention.

[0035] Preferably, when converting short-term memory into long-term memory, the corresponding short-term memory attention is used as the long-term memory attention.

[0036] The present invention also provides a system for realizing long-term memory of a digital human based on a mental model and a large language model, comprising:

[0037] An acquisition module, used to acquire perception data through sensors;

[0038] A preprocessing module is used to preprocess the sensing data to obtain preprocessed data;

[0039] A short-term memory building module is used to build short-term memory based on preprocessed data;

[0040] Long-term memory building blocks for accessing long-term memories from short-term memory;

[0041] A query module, used to query short-term memory and long-term memory according to the query content to obtain the query result;

[0042] The management module is used to manage long-term memory according to the query results.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. Improve memory management efficiency: Summarizing and classifying memories through a large language model can effectively reduce redundant information and improve memory management efficiency.

[0045] 2. Enhance memory query capabilities: Using a large language model to build multiple roles to associate and process information can dig deeper into memory content and improve the accuracy and comprehensiveness of memory queries.

[0046] 3. Optimize memory level management: Through attention management and memory level management, the storage level of memory can be dynamically adjusted to reduce unnecessary storage and computing overhead.

[0047] 4. Improve system stability: Run the memory sorting program regularly to ensure the accuracy and effectiveness of memory data and improve system stability and reliability.

[0048] 5. Provide personalized services: Through effective management and query of memory, it can provide users with more personalized and intelligent services and enhance user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0050] Figure 1 It is a flow chart of a method for realizing long-term memory of a digital human based on a mental model and a large language model according to the present invention. DETAILED DESCRIPTION

[0051] Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present invention, so the present invention is not limited to the specific implementation disclosed below.

[0052] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0053] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0054] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0055] like Figure 1As shown, the present invention provides a method for realizing long-term memory of a digital human based on a mental model and a large language model, comprising the following steps:

[0056] Acquire sensory data through sensors;

[0057] Preprocessing the sensed data to obtain preprocessed data;

[0058] Construct short-term memory based on preprocessed data;

[0059] Access long-term memory from short-term memory;

[0060] Query the short-term memory and the long-term memory according to the query content to obtain the query result;

[0061] Manage long-term memory based on query results.

[0062] Preferably, the sensing data is preprocessed to obtain preprocessed data, which specifically includes the following steps:

[0063] Convert the sensory data into text description information and store the text description information as the sensory_data field;

[0064] The sensor_data fields of the sensor data collected at the same time are grouped together and timestamps are added as preprocessed data.

[0065] Preferably, constructing short-term memory based on the preprocessed data specifically includes the following steps:

[0066] The preprocessed data is merged to obtain short-term memory.

[0067] Preferably, acquiring long-term memory from short-term memory specifically comprises the following steps:

[0068] All short-term memories under session IDs that have not added new data for more than two days are used as long-term memories.

[0069] Preferably, querying the short-term memory and the long-term memory according to the query content to obtain the query result specifically includes the following steps:

[0070] Build your character;

[0071] Get the query content;

[0072] Associating and processing the query content according to the role to obtain a label;

[0073] According to the label, query is performed in short-term memory and long-term memory to obtain the query result.

[0074] Preferably, managing long-term memory according to the query result specifically includes the following steps:

[0075] When the tag query successfully matches the short-term memory, the corresponding short-term memory attention is increased;

[0076] When the label query and the on-stage memory are matched successfully, the corresponding long-term memory attention is increased;

[0077] At regular intervals, reduce the focus on short-term memory and long-term memory;

[0078] When the long-term memory attention is lower than the attention threshold, the corresponding long-term memory is deleted.

[0079] Preferably, each time a query is successfully matched, the corresponding short-term memory attention or long-term memory attention is increased by one point;

[0080] Every other day, the short-term memory focus or long-term memory focus will decrease by one point.

[0081] Preferably, the method further comprises the following steps:

[0082] According to the long-term memory attention, the corresponding long-term memory indicator weight is obtained;

[0083] According to the long-term memory and the indicator weights, the short-term memory is scored to obtain the initial short-term memory attention.

[0084] Preferably, when converting short-term memory into long-term memory, the corresponding short-term memory attention is used as the long-term memory attention.

[0085] The present invention also provides a system for realizing long-term memory of a digital human based on a mental model and a large language model, comprising:

[0086] An acquisition module, used to acquire perception data through sensors;

[0087] A preprocessing module is used to preprocess the sensing data to obtain preprocessed data;

[0088] A short-term memory building module is used to build short-term memory based on preprocessed data;

[0089] Long-term memory building blocks for accessing long-term memories from short-term memory;

[0090] A query module, used to query short-term memory and long-term memory according to the query content to obtain the query result;

[0091] The management module is used to manage long-term memory according to the query results.

[0092] In order to better illustrate the technical effect of the present invention, the present invention provides the following specific embodiments to illustrate the above technical process:

[0093] Embodiment 1: A system for realizing long-term memory of digital human based on mental model and large language model, the system is divided into perception layer, short-term memory layer, long-term memory layer and working memory layer, such as Figure 1 As shown, the following steps are included:

[0094] S1, sensory input;

[0095] Through different types of sensors, such as vision, hearing, touch, etc., corresponding to video, sound, pressure and other sensor data, the information of the external environment is obtained and converted into text description information and stored as the sensor_data field. Since there may be multiple sensors in the same scene at the same time, the data collected at the same time is stored as a group as the data of the same conversation_id. Each data has a timestamp to identify the time when the data was collected.

[0096] S2, building short-term memory through language model;

[0097] Use a large language model (such as GPT) to process the input of the perception layer, and combine it with the relevant information in the long-term memory to form a short-term memory. First, for the perception data of the same group of conversations conversation_id, use a large language model (such as GPT) to merge the input of the perception layer to form a complete perception information as short-term memory. According to the complete perception information formed, combined with the relevant information in the long-term memory, the perception information is scored with a "0-10" attention through the large language model. There will be problems with doing this alone. GPT will not "mark the attention" according to AI's own unique knowledge and experience, and unique knowledge and experience should indeed affect the attention. For example, according to common sense, GPT will not pay attention to information such as "eat mango cake at a dinner party", but if the user is allergic to mangoes, then paying attention to something like "eat mango cake at a dinner party" is very meaningful. Therefore, the complete attention annotation will combine the associated related information in the long-term memory, and this information has attention in the long-term memory. Use this as the attention mark to score the short-term memory, and let GPT score the target information with a "0-10" attention according to the attention of the associated long-term memory information. In this way, the attention paid to new information can be determined based on the attention paid to existing information in experience.

[0098] S3, building long-term memory through memory organization;

[0099] The memory sorting program is always running. If no new data is added under the same conversation id for more than two days, all short-term memories under the conversation id will be used as long-term memories. All records in the conversations with the same conversation id (conversation_id) that have not added new data for more than two days in the short-term memory layer are continuously obtained and classified and sorted through the large language model. Memory sorting includes forgetting and awakening, multi-level storage, and memory addition, deletion, modification and query. The sorting content includes extracting memory type, memory label and other fields through the large language model. Memory types include semantic memory, episodic memory, procedural memory and other types.

[0100] Memory type: semantic memory stores general knowledge such as vocabulary, concepts, and rules. It is independent of time, place, and conditions, is relatively stable, and is not easily disturbed by external factors.

[0101] Memory type: Episodic (situational memory) stores specific information about personal experiences, events, situations, etc. It is dependent on time, place and conditions and is easily disturbed by external factors. The information is closely related to personal experience.

[0102] Memory type: procedural memory stores operational knowledge such as skills, habits, and rules. It is independent of time, place, and conditions, relatively stable, and not easily disturbed by external factors. It is usually less likely to change and can be exercised automatically without being aware of it, such as riding a bicycle, typing, etc. Memory tags are used to store memory tags, such as the source of the memory, the event to which it belongs, the scene of the memory, etc.

[0103] The attention of long-term memory is obtained by summing two variables: long-term attention and short-term attention. Long-term attention changes slowly and decays slowly; no matter what the reason, the increase in attention is the process of awakening, and the decrease in attention is the process of forgetting. The reason for the decrease in attention is relatively simple, that is, natural decay, and the increase in attention is related to the recall and use of information. For long-term attention, every time information is recalled and associated (effectively used), some long-term attention will be increased, and more short-term attention will be increased; recall and association will increase long-term attention by 1 point, short-term attention by 2 points, and automatically deduct 1 point every few days; short-term attention also has a source of awakening: single-point information in long-term memory is recalled and associated (used), which will increase the short-term attention of related information. We call this mechanism "regional awakening". Evidence of regional awakening can be found in humans. When we continue to think in a certain field, we can awaken the long-term memory of this field. The effect of this awakening is that it can greatly improve the thinking efficiency in this field.

[0104] The memory level is used to solve the problem of large amount of calculation for long-term memory content. Based on the attention level, all memories entering long-term memory have a default memory level of 1. As the attention level decreases, the memory level decreases by one when the attention level reaches 0. The maximum value of the memory level is configured according to the actual application scenario. When the attention level is 0 and the memory level is equal to the minimum value, the memory is deleted.

[0105] S4, query memory;

[0106] Query memory first uses a large language model (such as GPT) to build multiple roles (curious, answerer, inferer) to associate and process information, and extract fields such as beliefs, goals, intentions, memory tags, and problem solving motivation. The association function is implemented by defining roles and their functions: Curious: responsible for asking questions and exploring new information. Answerer: responsible for answering questions and providing explanations. Inferer: responsible for reasoning and predicting future situations. Use the GPT model to simulate role behavior and design specific prompts for each role to guide the GPT model to generate corresponding outputs. Integrate the role outputs more in-depth query data, and query short-term memory and long-term memory based on this data. Among them, the problem solving motivation is used to determine the hierarchical depth of querying long-term memory.

[0107] This embodiment implements the above content by using Python language, MongoDB storage, and a large language model (such as GPT):

[0108] 1. Data Storage

[0109] Use MongoDB to store both short-term and long-term memory.

[0110] Short-term memory is stored in the `ShortTermMemory` collection, and long-term memory is stored in the `LongTermMemory` collection.

[0111] Query memory is stored in the QueryMemory collection.

[0112] 2. Memory organization

[0113] The memory sorting program is always running, and all records in the conversations with the same conversation id (conversation_id) that have not had any new data for more than two days in the short-term memory layer are summarized and classified through the large language model.

[0114] Memory organization includes forgetting and awakening, multi-level storage, and adding, deleting, modifying and checking memories.

[0115] The organized content includes fields such as memory type, memory label, etc. extracted through a large language model.

[0116] 3. Memory query

[0117] Query memory first uses a large language model (such as GPT) to build multiple roles (curious, answerer, inferer) to associate and process information, and extract fields such as beliefs, goals, intentions, memory tags, and problem solving motivation.

[0118] Design specific prompts for each character to guide the GPT model to generate corresponding outputs.

[0119] Integrate roles to output more in-depth query data and query short-term and long-term memory based on this data.

[0120] 4. Memory level management

[0121] The memory level is used to solve the problem of large amount of calculation required to remember long-term content. Based on the degree of attention, the default memory level for all memories entering long-term memory is 1.

[0122] As the attention level decreases, when the attention level reaches 0, the memory level is reduced by one.

[0123] The maximum value of the memory level is configured according to the actual application scenario. When the attention level is 0 and the memory level is equal to the maximum value, the memory is deleted.

[0124] 5. Attention Management

[0125] The change in attention in long-term memory is obtained by summing two variables: long-term attention and short-term attention.

[0126] The higher the attention, the greater the weight of the data indicator used as a reference for GPT when the user queries data, so that GPT can pay more attention to the impact of current long-term memory on the query.

[0127] Long-term attention changes slowly and decays slowly; an increase in attention for whatever reason is a process of awakening, and a decrease in attention is a process of forgetting.

[0128] The reason for the decrease in attention is relatively simple, which is natural attenuation, while the increase in attention is related to the recall and use of information.

[0129] For long-term attention, every time information is recalled (used effectively), it increases some long-term attention and more short-term attention.

[0130] There is another source of arousal for short-term attention: when a single point of information in long-term memory is recalled (used), it will increase the short-term attention of related information. We call this mechanism "regional arousal".

[0131] 6. System Architecture

[0132] The system architecture includes data storage layer, memory organization layer, memory query layer, memory level management layer and attention management layer.

[0133] The data storage layer is responsible for storing short-term memory, long-term memory, and query memory.

[0134] The memory organization layer is responsible for summarizing and classifying memories.

[0135] The memory query layer is responsible for memory query using the large language model.

[0136] The memory hierarchy management layer is responsible for managing the memory hierarchy.

[0137] The attention management layer is responsible for managing the attention of memories.

[0138] 7. System Implementation

[0139] Use Python programming language to implement system functions.

[0140] Use MongoDB for data storage.

[0141] Use the GPT model for memory organization and query.

[0142] Run the memory sorting program regularly to ensure the accuracy and validity of the memory data.

[0143] Provides an API interface for external system calls to implement memory query and management functions.

[0144] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules, modules or units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units, modules or components can be combined or integrated into another device, or some features can be ignored or not executed.

[0145] The units may or may not be physically separated, and the components displayed as units may be one physical unit or multiple physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0146] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0147] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part, and / or installed from a removable medium. When the computer program is executed by the central processing unit (CPU), the above-mentioned functions defined in the method of the present invention are executed. It should be noted that the above-mentioned computer-readable medium of the present invention can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of an electrical, magnetic, optical, electromagnetic, infrared segment, or semiconductor, or any combination of the above.

[0148] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0149] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for realizing long-term memory of digital human based on mental model and large language model, characterized in that: The following steps are involved: Acquire sensory data through sensors; Preprocessing the sensed data to obtain preprocessed data; Construct short-term memory based on preprocessed data; Access long-term memory from short-term memory; Query the short-term memory and the long-term memory according to the query content to obtain the query result; Manage long-term memory based on query results.

2. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 1, characterized in that: The sensing data is preprocessed to obtain preprocessed data, which specifically includes the following steps: Convert the sensory data into text description information and store the text description information as the sensory_data field; The sensor_data fields of the sensor data collected at the same time are grouped together and timestamps are added as preprocessed data.

3. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 2, characterized in that: According to the preprocessed data, short-term memory is constructed, which includes the following steps: The preprocessed data is merged to obtain short-term memory.

4. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 3 is characterized in that: Acquiring long-term memory from short-term memory involves the following steps: The short-term memory under the session ID that has not added new data for more than two days is used as long-term memory.

5. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 4 is characterized in that: The short-term memory and the long-term memory are queried according to the query content to obtain the query result, which specifically includes the following steps: Build your character; Get the query content; Associating and processing the query content according to the role to obtain a label; According to the label, query is performed in short-term memory and long-term memory to obtain the query result.

6. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 5, characterized in that: Based on the query results, long-term memory is managed, including the following steps: When the tag query successfully matches the short-term memory, the corresponding short-term memory attention is increased; When the label query and the on-stage memory are matched successfully, the corresponding long-term memory attention is increased; At regular intervals, reduce the focus on short-term memory and long-term memory; When the long-term memory attention is lower than the attention threshold, the corresponding long-term memory is deleted.

7. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 6, characterized in that: Each time a query is successfully matched, the corresponding short-term memory attention or long-term memory attention is increased by one point; Every other day, the short-term memory focus or long-term memory focus will decrease by one point.

8. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 7, characterized in that: The following steps are also included: According to the long-term memory attention, the corresponding long-term memory indicator weight is obtained; According to the long-term memory and the indicator weights, the short-term memory is scored to obtain the initial short-term memory attention.

9. The method for realizing long-term memory of digital human based on mental model and large language model according to claim 8, characterized in that: When converting short-term memory into long-term memory, the corresponding short-term memory attention is used as the long-term memory attention.

10. A system for realizing long-term memory of digital human based on mental model and large language model, used for realizing the method for realizing long-term memory of digital human based on mental model and large language model as claimed in any one of claims 1 to 9, characterized in that: include: An acquisition module, used to acquire perception data through sensors; A preprocessing module is used to preprocess the sensing data to obtain preprocessed data; A short-term memory building module is used to build short-term memory based on preprocessed data; Long-term memory building blocks for accessing long-term memories from short-term memory; A query module, used to query short-term memory and long-term memory according to the query content to obtain the query result; The management module is used to manage long-term memory according to the query results.

Images