A hierarchical memory processing method and system of a personal AI assistant, a terminal, and a storage medium

Abstract

This invention relates to the field of intelligent voice data processing technology, and discloses a hierarchical memory processing method, system, terminal, and storage medium for a personal AI assistant. The method includes: acquiring and transcribing an audio stream and extracting a voiceprint feature stream; adaptively determining the conversation boundary based on the voiceprint feature stream, topic distribution changes, and silence pause duration to obtain contextual event units; classifying events into high, medium, and low value levels; constructing a three-layer memory structure containing instantaneous working memory, contextual event memory, and a semantic concept graph; anonymizing and projecting high and medium value events to obtain secure projection vectors and constructing a cloud-based nearest neighbor index; deconstructing user queries to obtain time, person, location, and semantic constraints; fusing and rearranging the three-way recall to obtain target event blocks, generating answers, and binding them to the original audio timestamps. This invention reduces data redundancy and retrieval noise, avoids confusion of facts across multiple subjects and times, balances cloud retrieval efficiency and privacy security, and provides a traceable chain of original audio evidence.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice data processing technology, and in particular to a hierarchical memory processing method, system, terminal, and computer-readable storage medium for a personal AI assistant. Background Technology

[0002] With the increasing prevalence of personal AI assistants and smart wearable devices, users expect systems to passively and continuously record daily life and work interactions, and accurately recall historical details or assist with pending tasks whenever needed. However, existing software algorithms and data processing solutions suffer from the following core technical shortcomings: Mechanical fixed slicing leads to semantic fragmentation and rampant noise in long-range retrieval: Existing systems typically segment text according to fixed durations or word counts. This method often abruptly cuts off a complete sentence or topic; simultaneously, if hundreds of thousands of words of trivial chatter from an entire day are fully saved and databaseed, the database will expand rapidly, recalling a large number of worthless and useless fragments during retrieval, resulting in a sharp drop in response accuracy. Flat storage structures cause serious confusion in memories of multiple people and across time: Existing patents mostly use flat table formats for storage. This structure cannot clarify complex causal relationships, producing serious factual illusions; the contradiction between privacy protection and semantic search usability during cloud retrieval: If the vector of the original text is directly uploaded to the cloud, there is a risk that the algorithm can reverse engineer it to reconstruct the original text; if traditional strong encryption algorithms are used, the ciphertext loses the geometric features of semantic similarity mathematically, making it impossible to perform fast similarity nearest neighbor searches in the cloud.

[0003] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0004] The main objective of this invention is to provide a hierarchical memory processing method, system, terminal, and computer-readable storage medium for personal AI assistants. This invention aims to solve the problems of semantic fragmentation and retrieval noise caused by fixed slicing, factual confusion caused by flat storage across multiple subjects and time, lack of memory decay, and contradiction between cloud privacy and semantic search, which lead to inaccurate long-term recall, factual confusion, and privacy leakage.

[0005] To achieve the above objectives, the present invention provides a hierarchical memory processing method for a personal AI assistant, the method comprising the following steps: A continuous audio stream is acquired, and speech transcription and voiceprint feature extraction are performed on the audio stream to obtain transcribed text and voiceprint feature stream; Based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, adaptive session boundary determination is performed to obtain context event units. The context event units are then classified by information value to obtain high-value events, medium-value events, and low-value events. Based on the high-value events and the medium-value events, a three-layer memory structure is constructed, which includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the raw, unsegmented streaming text, the contextual event memory is used to store the high-value events and the medium-value events, and the semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events. On the client side, entity occlusion and orthogonal transformation are performed on the high-value events and the medium-value events to obtain secure projection vectors, and the secure projection vectors are synchronized to the cloud to build a cloud-based nearest neighbor index library; Receive user query statements, deconstruct the user query statements to obtain time constraints, person constraints, location constraints and semantic constraints; Based on the time constraint, the person constraint, the location constraint, and the semantic constraint, spatiotemporal pruning is performed in the contextual event memory, graph relationship traversal is performed in the semantic concept graph, and encrypted vector matching is performed in the cloud nearest neighbor index database to obtain three-way recall results; The three recall results are merged and rearranged to obtain the target event block. A response is generated based on the target event block and the original audio timestamp is bound.

[0006] Optionally, the hierarchical memory processing method for the personal AI assistant, wherein the adaptive session boundary determination based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, to obtain contextual event units, and the information value classification of the contextual event units to obtain high-value events, medium-value events, and low-value events, specifically includes: Calculate the speaker feature distance between the preceding and following audio segments in the speaker feature stream to obtain the speaker change value; calculate the difference in the probability distribution of words within the preceding and following sliding windows in the transcribed text to obtain the topic divergence value; monitor the silence pause time at the end of the current sentence in the transcribed text to obtain the pause silence value. The normalized weighted sum of the speaker variation value, the topic divergence value, and the pause silence value is used to obtain the conversation boundary score. When the session boundary score exceeds the preset segmentation threshold, the current event is closed at the corresponding position to obtain the context event unit; Calculate the lexical information richness and key entity occurrence density of the contextual event unit, and classify the contextual event unit into high-value events, medium-value events and low-value events based on the lexical information richness and key entity occurrence density; Among them, the high-value events retain the full text, millisecond-level timestamps of each word, and speaker tags; the mid-value events extract the key facts and remove colloquial filler words; and the low-value events only record the time period and coarse classification tags, and are not written into the text retrieval index.

[0007] Optionally, the hierarchical memory processing method for the personal AI assistant, wherein constructing a three-layer memory structure based on the high-value events and the medium-value events specifically includes: The raw, unsegmented streaming text within a preset time period is written into a first-in-first-out circular sliding buffer to construct the instantaneous working memory. The high-value events and the medium-value events are written into the contextual event memory, and a unique event ID, a precise start and end time interval, environment and participant tags, a structured summary of the event, and a physical index pointer to the underlying audio encrypted storage block are generated for each event to construct the contextual event memory; Entity identification and relation extraction are performed on the high-value events and the medium-value events to obtain entities and relational connections connecting the entities, and the semantic concept graph is constructed. Each relational connection is established with a bidirectional tracing pointer pointing to the corresponding event in the contextual event memory.

[0008] Optionally, the hierarchical memory processing method for the personal AI assistant, wherein the construction of a three-layer memory structure based on the high-value events and the medium-value events further includes: Assign a dynamic activation value to each relation connection in the semantic concept graph; Each relationship is assigned a different half-life based on its importance level, so that temporary and trivial facts gradually fade over time and are moved to archived cold storage, while core facts containing agreements, commitments and deadlines remain highly active. When a historical fact is mentioned again in a subsequent new conversation or is hit by a user's query, a step boost is applied to the activation value of the historical fact and the decay clock is reset. When a newly occurring event conflicts with the historical facts, the timeliness is compared with the current activation value to establish a version overriding relationship and eliminate outdated facts.

[0009] Optionally, the hierarchical memory processing method for the personal AI assistant, wherein the step of performing entity occlusion and orthogonal transformation on the high-value events and the medium-value events on the client side to obtain a secure projection vector, and synchronizing the secure projection vector to the cloud to construct a cloud-based nearest neighbor index, specifically includes: On the client side, sensitive entities in the high-value events and medium-value events are identified, and sensitive names, accounts, amounts and private locations are replaced with anonymous hash placeholders to obtain desensitized text; The desensitized text is converted into an original high-dimensional semantic embedding vector, and the original high-dimensional semantic embedding vector is geometrically rotated using a client-private random orthogonal transformation matrix to obtain a rotated vector; Controlled differential noise is superimposed on the rotation vector to obtain the secure projection vector, and the secure projection vector is synchronized to the cloud. The secure projection vector is used as index data to construct the cloud-based nearest neighbor index library that supports similarity nearest neighbor search.

[0010] Optionally, the hierarchical memory processing method for the personal AI assistant, wherein receiving a user query statement and deconstructing the user query statement to obtain time constraints, person constraints, location constraints, and semantic constraints, specifically includes: Natural language parsing and named entity recognition are performed on the user query, and time range, people, location and core semantics are extracted from the recognition results; The time constraint is obtained based on the time range, the person constraint is obtained based on the person, the location constraint is obtained based on the location, and the semantic constraint is obtained based on the core semantics.

[0011] Optionally, the hierarchical memory processing method for the personal AI assistant, wherein the step of performing spatiotemporal pruning in the contextual event memory, traversing the semantic concept graph for graph relationships, and performing encrypted vector matching in the cloud-based nearest neighbor index database, based on the time constraint, the person constraint, the location constraint, and the semantic constraint, to obtain a three-way recall result, specifically includes: Based on the time constraints and the location constraints, the contextual event memory is filtered by time window and location to obtain the first recall result; Based on the character constraints, starting from the entities in the semantic concept graph, we expand outward along the relationship lines to search for related fact triples, and obtain the second recall result. Based on the semantic constraints, the user query statement is processed using the same entity masking and orthogonal transformation method as the secure projection vector to obtain a de-identified query vector. The de-identified query vector is then used to search for the most semantically similar event fragment in the cloud-based nearest neighbor index to obtain the third recall result.

[0012] Furthermore, to achieve the above objectives, the present invention also provides a hierarchical memory processing system for a personal AI assistant, wherein the hierarchical memory processing system for the personal AI assistant includes: The speech transcription and feature extraction module is used to acquire a continuously input audio stream, perform speech transcription and voiceprint feature extraction on the audio stream, and obtain transcribed text and voiceprint feature stream. The information value classification module is used to perform adaptive session boundary determination based on the voiceprint feature stream, the topic distribution change of the transcribed text, and the duration of silence pauses monitored from the audio stream, to obtain context event units, and to classify the information value of the context event units to obtain high-value events, medium-value events, and low-value events. A memory structure construction module is used to construct a three-layer memory structure based on the high-value events and the medium-value events. The three-layer memory structure includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the raw, unsegmented streaming text. The contextual event memory is used to store the high-value events and the medium-value events. The semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events. The geometric space transformation module is used to perform entity occlusion and orthogonal transformation on the high-value events and the medium-value events on the client side to obtain a safe projection vector, and synchronize the safe projection vector to the cloud to build a cloud nearest neighbor index library; The user intent deconstruction module is used to receive user query statements, deconstruct the user query statements to obtain time constraints, person constraints, location constraints and semantic constraints. The three-way parallel recall module is used to perform spatiotemporal pruning in the contextual event memory, graph relationship traversal in the semantic concept graph, and encrypted vector matching in the cloud nearest neighbor index database according to the time constraint, the person constraint, the location constraint, and the semantic constraint, respectively, to obtain the three-way recall results. The deterministic output module is used to fuse and rearrange the three recall results to obtain the target event block, generate a response based on the target event block, and bind the original audio timestamp.

[0013] In addition, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a hierarchical memory processing program for a personal AI assistant stored in the memory and executable on the processor, wherein when the hierarchical memory processing program for the personal AI assistant is executed by the processor, it implements the steps of the hierarchical memory processing method for the personal AI assistant as described above.

[0014] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a hierarchical memory processing program for a personal AI assistant, and when the hierarchical memory processing program for the personal AI assistant is executed by a processor, it implements the steps of the hierarchical memory processing method for a personal AI assistant as described above.

[0015] In this invention, an audio stream is acquired and its voiceprint feature stream is extracted through transcription. Adaptive session boundary determination is performed based on the voiceprint feature stream, topic distribution changes, and silence duration to obtain contextual event units. These units are then categorized by information value to obtain high, medium, and low-value events. A three-layer memory structure is constructed, comprising instantaneous working memory, contextual event memory, and a semantic concept graph. The contextual event memory stores high and medium-value events, while the semantic concept graph stores the structured fact network extracted from it. Entity masking and orthogonal transformation are performed on the high and medium-value events on the client side to obtain secure projection vectors, and a nearest neighbor index is simultaneously built in the cloud. User query statements are received and their intents are deconstructed to obtain time, person, location, and semantic constraints. Based on these constraints, spatiotemporal pruning, graph relationship traversal, and encrypted vector matching are performed in the contextual event memory, semantic concept graph, and cloud nearest neighbor index to obtain three-way recall results. These results are then fused and rearranged to obtain the target event block, generating an answer and binding it to the original audio timestamp. This invention reduces data redundancy and retrieval noise, avoids confusion of facts across multiple subjects and time periods, balances cloud retrieval efficiency and privacy security, and provides a traceable original evidence chain. Attached Figure Description

[0016] Figure 1 This is a flowchart of a preferred embodiment of the hierarchical memory processing method for the personal AI assistant of the present invention; Figure 2 This is a structural diagram of a preferred embodiment of the hierarchical memory processing system for a personal AI assistant of the present invention; Figure 3 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0017] This application provides a hierarchical memory processing method, system, and terminal for a personal AI assistant. To make the purpose, technical solution, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0018] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0020] The hierarchical memory processing method for a personal AI assistant described in a preferred embodiment of the present invention, such as... Figure 1 As shown, the hierarchical memory processing method of the personal AI assistant includes the following steps: Step S10: Obtain a continuously input audio stream, perform speech transcription and voiceprint feature extraction on the audio stream to obtain transcribed text and voiceprint feature stream.

[0021] It is understandable that personal AI assistants and smart wearable devices continuously collect audio throughout the day when passively and continuously recording daily life and work communication. Therefore, this step obtains a continuously input audio stream, rather than isolated audio segments uploaded in segments.

[0022] Furthermore, two processes are performed in parallel on the audio stream: speech transcription, which converts the speech content in the audio into text to obtain transcribed text; and voiceprint feature extraction, which extracts voice features that can characterize the speaker's identity from the audio to obtain a voiceprint feature stream. The transcribed text is a record of the conversation content in text form, and the voiceprint feature stream is feature data used to characterize different speakers.

[0023] Step S20: Based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, perform adaptive session boundary determination to obtain context event units, and classify the information value of the context event units to obtain high-value events, medium-value events, and low-value events.

[0024] Specifically, the speaker change value is obtained by calculating the speaker feature distance between the preceding and following audio segments in the speaker feature stream, calculating the difference in the probability distribution of words within the preceding and following sliding windows in the transcribed text, obtaining the topic divergence value, and monitoring the silence pause time at the end of the current sentence in the transcribed text. The normalized weighted sum of the speaker variation value, the topic divergence value, and the pause silence value is used to obtain the conversation boundary score. When the session boundary score exceeds the preset segmentation threshold, the current event is closed at the corresponding position to obtain the context event unit; Calculate the lexical information richness and key entity occurrence density of the contextual event unit, and classify the contextual event unit into high-value events, medium-value events and low-value events based on the lexical information richness and key entity occurrence density; Among them, the high-value events retain the full text, millisecond-level timestamps of each word, and speaker tags; the mid-value events extract the key facts and remove colloquial filler words; and the low-value events only record the time period and coarse classification tags, and are not written into the text retrieval index.

[0025] In this embodiment, adaptive conversation boundary determination is performed on continuously input transcribed text and voiceprint feature streams. The system uses a sliding window to monitor the transcribed text and voiceprint feature streams, determining whether a conversation has ended based on comprehensive changes in three dimensions. Dimension 1 is speaker change, which monitors the voiceprint feature distance between preceding and following audio segments to determine if the main speaker has changed, thus obtaining a speaker change value. Dimension 2 is topic content drift, which analyzes the difference in word probability distribution within the preceding and following windows, i.e., topic divergence, to determine if the discussion topic has substantially shifted, thus obtaining a topic divergence value. Dimension 3 is pause silence duration, which monitors whether the silence duration at the end of the current sentence exceeds the normal inter-speech pause threshold, thus obtaining a pause silence value. The changes in the above three dimensions are normalized and weighted to obtain a conversation boundary score. When the conversation boundary score exceeds a preset segmentation threshold, the system automatically closes the current event at that position, generating an independent contextual event unit, thereby avoiding the mechanical segmentation of complete semantics.

[0026] Information value grading is performed on each segmented contextual event unit. The system calculates the lexical information richness (i.e., semantic entropy) and key entity occurrence density of each event unit. Key entities include names, items, times, action words, etc. Based on lexical information richness and key entity occurrence density, the event unit is divided into three processing levels.

[0027] High-value events include explicit commitments, conclusions, or key tasks. For these events, the full text, including word-for-word text, millisecond-level timestamps for each word, and speaker tags, is retained. Medium-value events are ordinary discussions or low-density exchanges. For these events, a lightweight summarization algorithm is used to extract key facts, removing meaningless filler words and reducing storage usage by approximately 70%. Low-value events are purely ambient noise or minimal casual conversation. For these events, only the time period and a coarse classification label are recorded; they are not written to the text retrieval index.

[0028] Step S30: Based on the high-value events and the medium-value events, construct a three-layer memory structure, which includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the unsegmented raw streaming text, the contextual event memory is used to store the high-value events and the medium-value events, and the semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events.

[0029] Specifically, the raw, unsegmented streaming text within a preset time period is written into a first-in-first-out circular sliding buffer to construct the instantaneous working memory; The high-value events and the medium-value events are written into the contextual event memory, and a unique event ID, a precise start and end time interval, environment and participant tags, a structured summary of the event, and a physical index pointer to the underlying audio encrypted storage block are generated for each event to construct the contextual event memory; Entity identification and relation extraction are performed on the high-value events and the medium-value events to obtain entities and relational connections connecting the entities, and the semantic concept graph is constructed. Each relational connection is established with a bidirectional tracing pointer pointing to the corresponding event in the contextual event memory.

[0030] Understandably, the three-layer memory structure includes instantaneous working memory, contextual event memory, and semantic concept graph, and the three layers are maintained as a three-layer logical topology in memory and database.

[0031] The first layer is the transient working memory, which is constructed by writing the raw, unsegmented text from the most recent few minutes into a first-in-first-out circular sliding buffer. The transient working memory only stores the raw, unsegmented text from the most recent few minutes, which can be directly accessed by the currently ongoing real-time conversation and is not used for long-range retrieval.

[0032] The second layer is the contextual event memory. High-value and medium-value events are written into this contextual event memory to construct the contextual event memory. During writing, a unique event ID, a precise start and end time range, environment and participant tags, a structured event summary, and a physical index pointer to the underlying encrypted audio storage block are generated for each event. The physical index pointer is used to retrieve the corresponding original audio after a successful retrieval.

[0033] The third layer is a semantic concept graph. Entity identification and relation extraction are performed on the obtained high-value and mid-value events to obtain entities and the relational connections connecting them. Each relational connection establishes a bidirectional tracing pointer pointing to the corresponding event in the contextual event memory, thus constructing the semantic concept graph. The entities include people, projects, locations, and items. The relational connections represent relationships or actions and are accompanied by constraint attributes, including a deadline and current status. The bidirectional tracing pointer is used to trace back from the relational connections in the graph to the corresponding event in the contextual event memory, and then locates the original audio through the physical index pointer.

[0034] Furthermore, the construction of a three-layer memory structure based on the high-value events and the medium-value events further includes: Assign a dynamic activation value to each relation connection in the semantic concept graph; Each relationship is assigned a different half-life based on its importance level, so that temporary and trivial facts gradually fade over time and are moved to archived cold storage, while core facts containing agreements, commitments and deadlines remain highly active. When a historical fact is mentioned again in a subsequent new conversation or is hit by a user's query, a step boost is applied to the activation value of the historical fact and the decay clock is reset. When a newly occurring event conflicts with the historical facts, the timeliness is compared with the current activation value to establish a version overriding relationship and eliminate outdated facts.

[0035] In this embodiment, the constructed semantic concept graph is subjected to dynamic temporal decay and reinforcement evolution of memory, so that the factual relationships in the graph are no longer statically stored, but dynamically change with time and usage, in order to simulate the laws of human cognition.

[0036] Specifically, each relation connection in the semantic concept graph is assigned a dynamic activation value. This dynamic activation value characterizes the memory strength of the fact corresponding to that relation connection at the current moment; a higher value indicates that the fact is more easily retrieved, while a lower value indicates that the fact is closer to being forgotten. The dynamic activation value serves as the basis for subsequent decay, reinforcement, and conflict resolution calculations.

[0037] Different half-lives are set for each relationship connection based on its importance level. For temporary and trivial facts, a shorter half-life is set, allowing them to gradually fade over time and be automatically moved to archived cold storage once their activity level drops to a preset lower limit, thus avoiding significant redundancy in the database over time. For core facts containing agreements, commitments, and deadlines, an extremely long half-life or an extremely high anti-fading coefficient is set, ensuring they maintain high activity levels for a longer period and are not erased over time.

[0038] When a historical fact is mentioned again in a subsequent new conversation or is retrieved by a user query, a step-wise boost is applied to the activation value of that historical fact, and its decay clock is reset. This step-wise boost refers to directly adding a preset increment to the original activation value of the historical fact, rather than increasing it slowly proportionally. This continuously strengthens the memory of repeatedly mentioned or used facts, keeping them in a highly activated state that is easily retrieved.

[0039] When a new event conflicts with historical facts, the system automatically establishes a version overriding relationship by comparing the timeliness and current activation level of both events, thus eliminating outdated facts. For example, if a new event indicates that a meeting time has been changed from Wednesday to Friday, the system compares the timeliness and activation level of the new and old facts, sets the new fact as the valid version, marks the old fact as expired, and reduces its activation level. This ensures that the facts in the graph are always consistent with the latest situation, avoiding the need to recall outdated facts that have been refuted during retrieval.

[0040] Step S40: On the client side, perform entity occlusion and orthogonal transformation on the high-value events and the medium-value events to obtain a secure projection vector, and synchronize the secure projection vector to the cloud to build a cloud-based nearest neighbor index library.

[0041] Specifically, the client identifies sensitive entities in the high-value events and medium-value events, and replaces sensitive names, accounts, amounts, and private locations with anonymous hash placeholders to obtain desensitized text; The desensitized text is converted into an original high-dimensional semantic embedding vector, and the original high-dimensional semantic embedding vector is geometrically rotated using a client-private random orthogonal transformation matrix to obtain a rotated vector; Controlled differential noise is superimposed on the rotation vector to obtain the secure projection vector, and the secure projection vector is synchronized to the cloud. The secure projection vector is used as index data to construct the cloud-based nearest neighbor index library that supports similarity nearest neighbor search.

[0042] In this embodiment, data anonymization and projection processing are performed on the client side, ensuring data privacy protection before it leaves the device. The cloud only receives security vectors that cannot be reversed. First, sensitive entities in the high-value and medium-value events are identified on the client side. These sensitive entities include sensitive names, accounts, amounts, and private locations. After identification, these sensitive entities are replaced with anonymous hash placeholders to obtain the anonymized text. At this point, the anonymized text no longer contains specific content that directly identifies the user or trade secrets, but retains the semantic structure of the sentences.

[0043] Furthermore, the de-identified text is converted into an original high-dimensional semantic embedding vector X. The original high-dimensional semantic embedding vector X represents the semantic content of the de-identified text, but it is still in an unencrypted state. If it is directly uploaded to the cloud, there is still a risk of it being reverse-engineered.

[0044] The original high-dimensional semantic embedding vector X is geometrically rotated using a client-private random orthogonal transformation matrix P, resulting in a rotated vector P·X. According to the mathematical properties of orthogonal matrices, the inner product and included angles between the rotated vectors remain completely unchanged, i.e., distance and angle preservation is achieved. The random orthogonal transformation matrix P is stored only on the client side and is not uploaded to the cloud; the original vector cannot be reconstructed externally without the random orthogonal transformation matrix P.

[0045] Controlled differential noise η is superimposed on the rotation vector P·X to obtain the secure projection vector. The injection amount of the controlled differential noise η is controlled, so that the secure projection vector maintains comparable semantic similarity while further preventing the possibility of reconstructing the original semantic vector through mathematical inverse solution.

[0046] Finally, the secure projection vector is synchronized to the cloud. The cloud uses the secure projection vector as index data to build a cloud-based nearest neighbor index library that supports similarity-based nearest neighbor search. The cloud receives the transformed secure projection vector, which can directly calculate similarity using standard vector nearest neighbor algorithms. However, since there is no client-private random orthogonal transformation matrix P, third parties cannot mathematically solve for the original semantic vector, thus ensuring searchability on the cloud while achieving irreversible privacy protection.

[0047] Step S50: Receive the user query statement, deconstruct the user query statement to obtain time constraints, person constraints, location constraints and semantic constraints.

[0048] Specifically, the user query statement is parsed using natural language and named entity recognition, and the time range, people, locations and core semantics are extracted from the recognition results; The time constraint is obtained based on the time range, the person constraint is obtained based on the person, the location constraint is obtained based on the location, and the semantic constraint is obtained based on the core semantics.

[0049] Understandably, user queries are in the form of natural language questions. For example, a user might ask, "What was the final price of the pricing plan that Lao Zhang and I discussed at Starbucks last week?" This statement contains multiple constraints such as time, people, location, and core semantics, but it is expressed in a mixed form of natural language and cannot be directly used for retrieval. Therefore, it is necessary to deconstruct the intent of the query.

[0050] In this embodiment, the user query statement is first subjected to natural language parsing and named entity recognition. Natural language parsing is used to understand the syntactic structure and semantic roles of the user query statement, distinguishing components such as time adverbs, person subjects, place adverbs, and core predicates and objects. Named entity recognition is used to locate and extract entity words with specific meanings from the user query statement, including time words, person names, place names, and business object words.

[0051] Next, the time range, people, location, and core semantics are extracted from the user query. The time range refers to the time information used to limit the retrieval time window in the user query, such as "last week," "yesterday afternoon," or "last month." The people refer to the target individuals mentioned in the user query, such as "Old Zhang" or "Engineer Li." The location refers to the spatial location information mentioned in the user query, such as "Starbucks" or "meeting room." The core semantics refers to the target thing or topic pointed to by the user query, such as "quotation plan" or "interface document."

[0052] Finally, the time constraint is obtained based on the time range, the person constraint is obtained based on the person, the location constraint is obtained based on the location, and the semantic constraint is obtained based on the core semantics. The time constraint, person constraint, location constraint, and semantic constraint are used as filtering and matching conditions for different paths in subsequent retrieval steps. Specifically, the time constraint and location constraint are used for spatiotemporal pruning in contextual event memory, the person constraint is used for graph relationship traversal in the semantic concept graph, and the semantic constraint is used for encrypted vector matching in the cloud-based nearest neighbor index.

[0053] Step S60: Based on the time constraint, the person constraint, the location constraint, and the semantic constraint, perform spatiotemporal pruning in the contextual event memory, perform graph relationship traversal in the semantic concept graph, and perform encrypted vector matching in the cloud nearest neighbor index database to obtain three-way recall results.

[0054] Specifically, based on the time constraint and the location constraint, the contextual event memory is filtered by time window and location to obtain the first recall result; Based on the character constraints, starting from the entities in the semantic concept graph, we expand outward along the relationship lines to search for related fact triples, and obtain the second recall result. Based on the semantic constraints, the user query statement is processed using the same entity masking and orthogonal transformation method as the secure projection vector to obtain a de-identified query vector. The de-identified query vector is then used to search for the most semantically similar event fragment in the cloud-based nearest neighbor index to obtain the third recall result.

[0055] In this embodiment, three parallel recalls are performed based on the above constraints (let the three paths be path A, path B, and path C, respectively). The three recalls are not simply superimposed, but rather cross-validate the same target event from the temporal and spatial dimensions, the graph relationship dimension, and the semantic similarity dimension, respectively. Then, the recall results of the three paths are merged and rearranged to finally generate an answer with the original anchor.

[0056] Path A involves spatiotemporal pruning. Specifically, based on the time and location constraints, filtering is performed on the contextual event memory according to time windows and location. The contextual event memory stores the precise start and end time intervals, environment, and participant tags for each event. Therefore, events outside the target time window and irrelevant locations can be directly excluded, yielding the first recall result. For example, if a user asks, "The pricing plan that Lao Zhang and I discussed at Starbucks last week," Path A uses last week as the time window and Starbucks as the location condition to quickly filter out events that did not occur within that time window or at that location, retaining only events that simultaneously meet the time and location conditions as the first recall result. The purpose of Path A is to quickly narrow the search scope, avoid semantic calculations on all historical events, and reduce long-range search noise.

[0057] Path B involves traversing the semantic concept graph. Specifically, based on the person constraint, it starts from entities in the semantic concept graph and expands outward along the relationship lines to find associated fact triples, obtaining the second recall result. Entities in the semantic concept graph include people, projects, locations, and items. The relationship lines represent relationships or actions with attached constraint attributes, including deadlines and current states. Path B starts from the entity corresponding to the person constraint and expands outward along the relationship lines, accurately capturing structured facts such as promises, conclusions, and pending tasks related to the target person. For example, starting from the entity "Old Zhang," it searches for associated fact triples along relationship lines such as "promised delivery," "participated in discussion," and "responsible for quoting," and then combines this with the project entity and the quotation scheme entity to obtain the second recall result. The purpose of Path B is to avoid missing key facts based solely on semantic similarity, especially suitable for scenarios where the person in the user query is clear but the specific description is vague, preventing confusion between "what Zhang San said" and "what Li Si said."

[0058] Path C involves encrypted vector matching. Specifically, based on the semantic constraints, the user query is processed using the same entity masking and orthogonal transformation methods as the secure projection vector to obtain a de-identified query vector. This de-identified query vector is then used to search for the semantically most similar event fragment in the cloud-based nearest neighbor index, yielding a third recall result. The cloud-based nearest neighbor index stores secure projection vectors synchronized to the cloud. The cloud only uses the de-identified query vector and the secure projection vector for similarity calculation, without knowing the plaintext of the user query or being able to decipher the original semantics in the index. The role of Path C is to recall the semantically closest text fragments, compensating for semantically relevant events that Path A and Path B might miss due to overly strict time, location, and person criteria.

[0059] Step S70: The three recall results are fused and rearranged to obtain the target event block. A response is generated based on the target event block and the original audio timestamp is bound.

[0060] In this embodiment, the first, second, and third recall results are subjected to a cross-attention secondary reordering. The cross-attention reordering model integrates the three scores, fuses and refines the three recall results, and accurately identifies the most core target event block to obtain the target event block. This cross-attention secondary reordering simultaneously considers temporal-spatial matching degree, graph relationship matching degree, and semantic similarity scores to avoid false recalls caused by single-dimensional scoring.

[0061] For example, path A may recall multiple events at the same time and place, path B may recall multiple facts related to Lao Zhang, and path C may recall semantically similar but not the same event fragments. After cross-attention reordering cross-compares the results of the three paths, the unique target event block that simultaneously satisfies the conditions of time, place, person, and semantics is locked.

[0062] Furthermore, the millisecond-level start and end timestamps of the original recording are read from the metadata of the target event block. Since the target event block already contains a physical index pointer pointing to the underlying encrypted audio storage block when constructing the contextual event memory, the corresponding original recording start and end times can be directly read. A conclusive answer is generated based on the target event block and bound to a clickable original audio playback anchor. Users can click the original audio playback anchor to listen back to the live dialogue, thereby resolving the illusion problem of large-scale question-and-answer models and providing a traceable and definitive chain of evidence for business communication, work commitments, and important decisions.

[0063] The beneficial effects of this invention are as follows: (1) Significantly reduce data redundancy and long-range retrieval distortion: This invention uses an adaptive slicing mechanism with dual acoustic and semantic gating to close events at natural conversation boundaries, filtering out more than 65% of meaningless redundant records from the source. The recall accuracy of long-range historical recall is improved by more than 35% compared with traditional fixed block segmentation.

[0064] (2) Completely solve the confusion of facts involving multiple subjects and long time sequence: The three-layer hierarchical memory network of the present invention completely decouples instantaneous dialogue, specific events and abstract entity concepts in data structure. Combined with the evolution mechanism of adaptive fading and reinforcement over time, it eliminates the "misattribution" of facts across time and across personnel.

[0065] (3) Balancing high-speed cloud retrieval with irreversible data security: The orthogonal geometric projection technology of this invention enables the ciphertext vector to maintain comparable similarity while mathematically blocking the path to reverse deciphering the plaintext, thus balancing cloud retrieval performance with privacy security.

[0066] (4) Trustworthy and deterministic evidence chain backtracking: In this invention, all answers are directly bound to the physical timestamp interval of the underlying original audio, supporting one-click original sound verification, providing a solid traceable guarantee for business communication, work commitments and important decisions.

[0067] Furthermore, such as Figure 2 As shown, based on the above-mentioned hierarchical memory processing method for personal AI assistants, the present invention also provides a hierarchical memory processing system for personal AI assistants, wherein the hierarchical memory processing system for personal AI assistants includes: The speech transcription and feature extraction module 51 is used to acquire a continuously input audio stream, perform speech transcription and voiceprint feature extraction on the audio stream, and obtain transcribed text and voiceprint feature stream. The information value classification module 52 is used to perform adaptive session boundary determination based on the voiceprint feature stream, the topic distribution change of the transcribed text, and the duration of silence pauses monitored from the audio stream, to obtain context event units, and to classify the information value of the context event units to obtain high-value events, medium-value events, and low-value events. The memory structure construction module 53 is used to construct a three-layer memory structure based on the high-value events and the medium-value events. The three-layer memory structure includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the raw, unsegmented streaming text. The contextual event memory is used to store the high-value events and the medium-value events. The semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events. The geometric space transformation module 54 is used to perform entity occlusion and orthogonal transformation on the high-value events and the medium-value events on the client side to obtain a safe projection vector, and synchronize the safe projection vector to the cloud to build a cloud nearest neighbor index library; User intent deconstruction module 55 is used to receive user query statements, deconstruct the user query statements to obtain time constraints, person constraints, location constraints and semantic constraints. The three-way parallel recall module 56 is used to perform spatiotemporal pruning in the contextual event memory, graph relationship traversal in the semantic concept graph, and encrypted vector matching in the cloud nearest neighbor index database according to the time constraint, the person constraint, the location constraint, and the semantic constraint, respectively, to obtain the three-way recall results. The deterministic output module 57 is used to fuse and rearrange the three recall results to obtain a target event block, generate a response based on the target event block, and bind the original audio timestamp.

[0068] Furthermore, such as Figure 3 As shown, based on the hierarchical memory processing method and system for the personal AI assistant described above, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 3 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0069] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard drive or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard drive, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a hierarchical memory processing program 40 for a personal AI assistant, which can be executed by the processor 10 to implement the hierarchical memory processing method for the personal AI assistant in this application.

[0070] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the hierarchical memory processing method of the personal AI assistant.

[0071] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0072] In one embodiment, when the processor 10 executes the hierarchical memory processing program 40 for the personal AI assistant in the memory 20, the following steps are performed: A continuous audio stream is acquired, and speech transcription and voiceprint feature extraction are performed on the audio stream to obtain transcribed text and voiceprint feature stream; Based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, adaptive session boundary determination is performed to obtain context event units. The context event units are then classified by information value to obtain high-value events, medium-value events, and low-value events. Based on the high-value events and the medium-value events, a three-layer memory structure is constructed, which includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the raw, unsegmented streaming text, the contextual event memory is used to store the high-value events and the medium-value events, and the semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events. On the client side, entity occlusion and orthogonal transformation are performed on the high-value events and the medium-value events to obtain secure projection vectors, and the secure projection vectors are synchronized to the cloud to build a cloud-based nearest neighbor index library; Receive user query statements, deconstruct the user query statements to obtain time constraints, person constraints, location constraints and semantic constraints; Based on the time constraint, the person constraint, the location constraint, and the semantic constraint, spatiotemporal pruning is performed in the contextual event memory, graph relationship traversal is performed in the semantic concept graph, and encrypted vector matching is performed in the cloud nearest neighbor index database to obtain three-way recall results; The three recall results are merged and rearranged to obtain the target event block. A response is generated based on the target event block and the original audio timestamp is bound.

[0073] Specifically, the adaptive session boundary determination based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, to obtain contextual event units, and the information value classification of the contextual event units to obtain high-value events, medium-value events, and low-value events, includes: Calculate the speaker feature distance between the preceding and following audio segments in the speaker feature stream to obtain the speaker change value; calculate the difference in the probability distribution of words within the preceding and following sliding windows in the transcribed text to obtain the topic divergence value; monitor the silence pause time at the end of the current sentence in the transcribed text to obtain the pause silence value. The normalized weighted sum of the speaker variation value, the topic divergence value, and the pause silence value is used to obtain the conversation boundary score. When the session boundary score exceeds the preset segmentation threshold, the current event is closed at the corresponding position to obtain the context event unit; Calculate the lexical information richness and key entity occurrence density of the contextual event unit, and classify the contextual event unit into high-value events, medium-value events and low-value events based on the lexical information richness and key entity occurrence density; Among them, the high-value events retain the full text, millisecond-level timestamps of each word, and speaker tags; the mid-value events extract the key facts and remove colloquial filler words; and the low-value events only record the time period and coarse classification tags, and are not written into the text retrieval index.

[0074] The construction of a three-layer memory structure based on the high-value events and the medium-value events specifically includes: The raw, unsegmented streaming text within a preset time period is written into a first-in-first-out circular sliding buffer to construct the instantaneous working memory. The high-value events and the medium-value events are written into the contextual event memory, and a unique event ID, a precise start and end time interval, environment and participant tags, a structured summary of the event, and a physical index pointer to the underlying audio encrypted storage block are generated for each event to construct the contextual event memory; Entity identification and relation extraction are performed on the high-value events and the medium-value events to obtain entities and relational connections connecting the entities, and the semantic concept graph is constructed. Each relational connection is established with a bidirectional tracing pointer pointing to the corresponding event in the contextual event memory.

[0075] The construction of a three-layer memory structure based on the high-value events and the medium-value events further includes: Assign a dynamic activation value to each relation connection in the semantic concept graph; Each relationship is assigned a different half-life based on its importance level, so that temporary and trivial facts gradually fade over time and are moved to archived cold storage, while core facts containing agreements, commitments and deadlines remain highly active. When a historical fact is mentioned again in a subsequent new conversation or is hit by a user's query, a step boost is applied to the activation value of the historical fact and the decay clock is reset. When a newly occurring event conflicts with the historical facts, the timeliness is compared with the current activation value to establish a version overriding relationship and eliminate outdated facts.

[0076] Specifically, the step of performing entity occlusion and orthogonal transformation on the high-value and medium-value events on the client side to obtain a secure projection vector, and synchronizing the secure projection vector to the cloud to build a cloud-based nearest neighbor index library, includes: On the client side, sensitive entities in the high-value events and medium-value events are identified, and sensitive names, accounts, amounts and private locations are replaced with anonymous hash placeholders to obtain desensitized text; The desensitized text is converted into an original high-dimensional semantic embedding vector, and the original high-dimensional semantic embedding vector is geometrically rotated using a client-private random orthogonal transformation matrix to obtain a rotated vector; Controlled differential noise is superimposed on the rotation vector to obtain the secure projection vector, and the secure projection vector is synchronized to the cloud. The secure projection vector is used as index data to construct the cloud-based nearest neighbor index library that supports similarity nearest neighbor search.

[0077] Specifically, the process of receiving a user query statement and deconstructing the user query statement to obtain time constraints, person constraints, location constraints, and semantic constraints includes: Natural language parsing and named entity recognition are performed on the user query, and time range, people, location and core semantics are extracted from the recognition results; The time constraint is obtained based on the time range, the person constraint is obtained based on the person, the location constraint is obtained based on the location, and the semantic constraint is obtained based on the core semantics.

[0078] Specifically, based on the time constraint, the person constraint, the location constraint, and the semantic constraint, spatiotemporal pruning is performed in the contextual event memory, graph relationship traversal is performed in the semantic concept graph, and encrypted vector matching is performed in the cloud-based nearest neighbor index to obtain three-way recall results, including: Based on the time constraints and the location constraints, the contextual event memory is filtered by time window and location to obtain the first recall result; Based on the character constraints, starting from the entities in the semantic concept graph, we expand outward along the relationship lines to search for related fact triples, and obtain the second recall result. Based on the semantic constraints, the user query statement is processed using the same entity masking and orthogonal transformation method as the secure projection vector to obtain a de-identified query vector. The de-identified query vector is then used to search for the most semantically similar event fragment in the cloud-based nearest neighbor index to obtain the third recall result.

[0079] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a hierarchical memory processing program for a personal AI assistant, and when the hierarchical memory processing program for the personal AI assistant is executed by a processor, it implements the steps of the hierarchical memory processing method for a personal AI assistant as described above.

[0080] In summary, this invention provides a hierarchical memory processing method, system, terminal, and storage medium for a personal AI assistant. The method includes: acquiring an audio stream and transcribing it to extract voiceprint feature streams; performing adaptive session boundary determination based on voiceprint feature streams, topic distribution changes, and silence duration to obtain contextual event units; classifying the information value of contextual event units to obtain high, medium, and low value events; constructing a three-layer memory structure containing instantaneous working memory, contextual event memory, and semantic concept graph, where contextual event memory stores high and medium value events, and the semantic concept graph stores the structured fact network extracted from it; performing entity masking and orthogonal transformation on high and medium value events on the client side to obtain secure projection vectors and simultaneously constructing a nearest neighbor index in the cloud; receiving user query statements and deconstructing intents to obtain time, person, location, and semantic constraints; performing spatiotemporal pruning, graph relationship traversal, and encrypted vector matching based on each constraint in contextual event memory, semantic concept graph, and cloud nearest neighbor index to obtain three-way recall results, fusing and rearranging them to obtain target event blocks, generating answers, and binding the original audio timestamps. This invention reduces data redundancy and retrieval noise, avoids confusion of facts across multiple subjects and time periods, balances cloud retrieval efficiency and privacy security, and provides a traceable original evidence chain.

[0081] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0082] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0083] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A hierarchical memory processing method for a personal AI assistant, characterized in that, The hierarchical memory processing method for the personal AI assistant includes: A continuous audio stream is acquired, and speech transcription and voiceprint feature extraction are performed on the audio stream to obtain transcribed text and voiceprint feature stream; Based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, adaptive session boundary determination is performed to obtain context event units. The context event units are then classified by information value to obtain high-value events, medium-value events, and low-value events. Based on the high-value events and the medium-value events, a three-layer memory structure is constructed, which includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the raw, unsegmented streaming text, the contextual event memory is used to store the high-value events and the medium-value events, and the semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events. On the client side, entity occlusion and orthogonal transformation are performed on the high-value events and the medium-value events to obtain secure projection vectors, and the secure projection vectors are synchronized to the cloud to build a cloud-based nearest neighbor index library; Receive user query statements, deconstruct the user query statements to obtain time constraints, person constraints, location constraints and semantic constraints; Based on the time constraint, the person constraint, the location constraint, and the semantic constraint, spatiotemporal pruning is performed in the contextual event memory, graph relationship traversal is performed in the semantic concept graph, and encrypted vector matching is performed in the cloud nearest neighbor index database to obtain three-way recall results; The three recall results are merged and rearranged to obtain the target event block. A response is generated based on the target event block and the original audio timestamp is bound.

2. The hierarchical memory processing method for a personal AI assistant according to claim 1, characterized in that, The process involves adaptively determining conversation boundaries based on the voiceprint feature stream, the topic distribution changes of the transcribed text, and the duration of silence pauses monitored from the audio stream, to obtain contextual event units. These contextual event units are then classified by information value into high-value events, medium-value events, and low-value events. Specifically, this includes: Calculate the speaker feature distance between the preceding and following audio segments in the speaker feature stream to obtain the speaker change value; calculate the difference in the probability distribution of words within the preceding and following sliding windows in the transcribed text to obtain the topic divergence value; monitor the silence pause time at the end of the current sentence in the transcribed text to obtain the pause silence value. The normalized weighted sum of the speaker variation value, the topic divergence value, and the pause silence value is used to obtain the conversation boundary score. When the session boundary score exceeds a preset segmentation threshold, the current event is closed at the corresponding position to obtain the context event unit; Calculate the lexical information richness and key entity occurrence density of the contextual event unit, and classify the contextual event unit into high-value events, medium-value events and low-value events based on the lexical information richness and key entity occurrence density; Among them, the high-value events retain the full text, millisecond-level timestamps of each word, and speaker tags; the mid-value events extract the key facts and remove colloquial filler words; and the low-value events only record the time period and coarse classification tags, and are not written into the text retrieval index.

3. The hierarchical memory processing method for a personal AI assistant according to claim 1, characterized in that, The construction of a three-layer memory structure based on the high-value events and the medium-value events specifically includes: The raw, unsegmented streaming text within a preset time period is written into a first-in-first-out circular sliding buffer to construct the instantaneous working memory. The high-value events and the medium-value events are written into the contextual event memory, and a unique event ID, a precise start and end time interval, environment and participant tags, a structured summary of the event, and a physical index pointer to the underlying audio encrypted storage block are generated for each event to construct the contextual event memory; Entity identification and relation extraction are performed on the high-value events and the medium-value events to obtain entities and relational connections connecting the entities, and the semantic concept graph is constructed. Each relational connection is established with a bidirectional tracing pointer pointing to the corresponding event in the contextual event memory.

4. The hierarchical memory processing method for a personal AI assistant according to claim 3, characterized in that, The process of constructing a three-layer memory structure based on the high-value events and the medium-value events further includes: Assign a dynamic activation value to each relation connection in the semantic concept graph; Each relationship is assigned a different half-life based on its importance level, so that temporary and trivial facts gradually fade over time and are moved to archived cold storage, while core facts containing agreements, commitments and deadlines remain highly active. When a historical fact is mentioned again in a subsequent new conversation or is hit by a user's query, a step boost is applied to the activation value of the historical fact and the decay clock is reset. When a newly occurring event conflicts with the historical facts, the timeliness is compared with the current activation value to establish a version overriding relationship and eliminate outdated facts.

5. The hierarchical memory processing method for a personal AI assistant according to claim 1, characterized in that, The step of performing entity occlusion and orthogonal transformation on the high-value and medium-value events on the client side to obtain a secure projection vector, and synchronizing the secure projection vector to the cloud to build a cloud-based nearest neighbor index library, specifically includes: On the client side, sensitive entities in the high-value events and medium-value events are identified, and sensitive names, accounts, amounts and private locations are replaced with anonymous hash placeholders to obtain desensitized text; The desensitized text is converted into an original high-dimensional semantic embedding vector, and the original high-dimensional semantic embedding vector is geometrically rotated using a client-private random orthogonal transformation matrix to obtain a rotated vector; Controlled differential noise is superimposed on the rotation vector to obtain the secure projection vector, and the secure projection vector is synchronized to the cloud. The secure projection vector is used as index data to construct the cloud-based nearest neighbor index library that supports similarity nearest neighbor search.

6. The hierarchical memory processing method for a personal AI assistant according to claim 1, characterized in that, The process of receiving a user query statement involves deconstructing the query statement to obtain time constraints, person constraints, location constraints, and semantic constraints, specifically including: Natural language parsing and named entity recognition are performed on the user query, and time range, people, location and core semantics are extracted from the recognition results; The time constraint is obtained based on the time range, the person constraint is obtained based on the person, the location constraint is obtained based on the location, and the semantic constraint is obtained based on the core semantics.

7. The hierarchical memory processing method for a personal AI assistant according to claim 3, characterized in that, Based on the time constraint, the person constraint, the location constraint, and the semantic constraint, the process involves spatiotemporal pruning in the contextual event memory, graph relationship traversal in the semantic concept graph, and encrypted vector matching in the cloud-based nearest neighbor index, respectively, to obtain three-way recall results, specifically including: Based on the time constraints and the location constraints, the contextual event memory is filtered by time window and location to obtain the first recall result; Based on the character constraints, starting from the entities in the semantic concept graph, we expand outward along the relationship lines to search for related fact triples, and obtain the second recall result. Based on the semantic constraints, the user query statement is processed using the same entity masking and orthogonal transformation method as the secure projection vector to obtain a de-identified query vector. The de-identified query vector is then used to search for the most semantically similar event fragment in the cloud-based nearest neighbor index to obtain the third recall result.

8. A hierarchical memory processing system for a personal AI assistant, characterized in that, The hierarchical memory processing system of the personal AI assistant includes: The speech transcription and feature extraction module is used to acquire a continuously input audio stream, perform speech transcription and voiceprint feature extraction on the audio stream, and obtain transcribed text and voiceprint feature stream. The information value classification module is used to perform adaptive session boundary determination based on the voiceprint feature stream, the topic distribution change of the transcribed text, and the duration of silence pauses monitored from the audio stream, to obtain context event units, and to classify the information value of the context event units to obtain high-value events, medium-value events, and low-value events. A memory structure construction module is used to construct a three-layer memory structure based on the high-value events and the medium-value events. The three-layer memory structure includes transient working memory, contextual event memory, and semantic concept graph. The transient working memory is used to store the raw, unsegmented streaming text. The contextual event memory is used to store the high-value events and the medium-value events. The semantic concept graph is used to store the structured fact network extracted from the high-value events and the medium-value events. The geometric space transformation module is used to perform entity occlusion and orthogonal transformation on the high-value events and the medium-value events on the client side to obtain a safe projection vector, and synchronize the safe projection vector to the cloud to build a cloud nearest neighbor index library; The user intent deconstruction module is used to receive user query statements, deconstruct the user query statements to obtain time constraints, person constraints, location constraints and semantic constraints. The three-way parallel recall module is used to perform spatiotemporal pruning in the contextual event memory, graph relationship traversal in the semantic concept graph, and encrypted vector matching in the cloud nearest neighbor index database according to the time constraint, the person constraint, the location constraint, and the semantic constraint, respectively, to obtain the three-way recall results. The deterministic output module is used to fuse and rearrange the three recall results to obtain the target event block, generate a response based on the target event block, and bind the original audio timestamp.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a hierarchical memory processing program for a personal AI assistant stored in the memory and executable on the processor. When the hierarchical memory processing program for the personal AI assistant is executed by the processor, it implements the steps of the hierarchical memory processing method for a personal AI assistant as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a hierarchical memory processing program for a personal AI assistant, which, when executed by a processor, implements the steps of the hierarchical memory processing method for a personal AI assistant as described in any one of claims 1-7.

Images