Memory anchoring system and method based on dialogue stream

Abstract

The invention discloses a memory anchoring system and method based on dialogue streams, and the system comprises a multi-mode input interface which is used for receiving natural dialogue streams; the progressive emotion mining device is used for carrying out progressive analysis on the natural dialogue flow and dynamically generating an emotion label; the emotion tree memory is used for dividing a storage path according to the emotion label so as to realize end-side hierarchical storage memory; the passive wake-up engine is used for performing passive wake-up detection based on the natural dialogue stream to distinguish a corresponding scene so as to judge whether historical memory association is triggered or not; the natural dialogue generator is used for performing intelligent calling of historical memory association and outputting a response if the natural dialogue generator is triggered; and if not, recording and storing a new entity, and generating a corresponding natural response. According to the method, memory continuity can be realized and memory depth faults can be eliminated through progressive emotion mining and tree-shaped storage according to emotions; historical memories are associated and fused into natural dialogue streams, so that the mechanical interaction feeling is eliminated.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence dialogue technology, specifically to a memory anchoring system and method based on dialogue flow. Background Technology

[0002] Current mainstream dialogue systems (such as Chat GPT) use a fixed-length context window (typically 128K tokens) to maintain dialogue coherence in the following ways:

[0003] Short-term memory: caching the dialogue text in memory;

[0004] Keyword matching: Retrieves historical information from a cloud database;

[0005] Proactively ask: Directly request users to repeat key information.

[0006] However, the above method has the following shortcomings:

[0007] The lack of a hierarchical storage interface in the memory management module causes gaps in memory depth, resulting in a high rate of historical forgetting after a conversation exceeds 128K tokens.

[0008] The independent query module is separated from the dialogue flow, making the interaction feel mechanical.

[0009] The retrieval architecture relies on remote servers, resulting in excessively high energy costs. Summary of the Invention

[0010] In view of the technical deficiencies mentioned in the background art, the purpose of this invention is to provide a memory anchoring system and method based on dialogue flow, which aims to at least partially solve one of the technical problems in the related art.

[0011] To achieve the above objectives, in a first aspect, embodiments of the present invention provide a dialogue-stream-based memory anchoring system applied to a smart terminal, the system comprising:

[0012] A multimodal input interface for receiving natural dialogue streams;

[0013] A progressive sentiment miner is used to progressively analyze the natural dialogue stream and dynamically generate sentiment tags.

[0014] An emotion tree-structured memory is used to divide storage paths according to the emotion tags in order to achieve hierarchical storage memory on the end side; wherein, the emotion tree-structured memory is constructed with tree-structured storage nodes, and the physical isolation storage of multi-dimensional features is achieved through the tree-structured storage nodes;

[0015] A passive wake-up engine is used to perform passive wake-up detection based on the natural dialogue flow, so as to distinguish the corresponding scenario and thus determine whether the historical memory association is triggered; wherein, the scenario includes instant communication and historical memory.

[0016] A natural dialogue generator for:

[0017] If triggered, the system will intelligently retrieve historical memories and output a response.

[0018] If not triggered, a new entity is recorded and stored, and a corresponding natural response is generated.

[0019] As a specific implementation of this application, the progressive analysis, which dynamically generates emotion tags, specifically includes:

[0020] The process involves shallow feature extraction, mid-level semantic parsing, and deep sentiment fusion. A sentiment value is calculated based on a preset weight adaptive rule, and a corresponding sentiment label is generated based on the magnitude of the sentiment value. The sentiment label includes strong events and weak events.

[0021] As a specific implementation of this application, storage paths are divided according to the emotion tags to achieve hierarchical storage and memory on the client side, specifically including:

[0022] Strong events: Store the raw data in the "emotion root node to strong emotion area" according to the four-level structure of "time branch to scene leaf to entity leaf to feature node";

[0023] Weak events: compressed into 128-dimensional feature vectors and stored in the "emotion root node to weak emotion area to feature vector pool".

[0024] As one specific implementation of this application, triggering historical memory association includes:

[0025] Keyword trigger: Preset keywords are detected, and related events within ± preset years are retrieved based on the current sentiment value.

[0026] Silence trigger: When the current conversation is paused for more than a set time, a light scan of multiple recent topic keywords is performed; and when related, priority is given to matching those with high emotional similarity and core entities.

[0027] In a preferred implementation of this application, the system further includes an anti-deviation module, which is used for:

[0028] If historical memory is triggered, the original target anchor point will be automatically activated, anti-deviation monitoring will be started, and a natural response will be generated.

[0029] If it is a new topic for instant communication, the mode is determined first; the mode includes work mode and casual chat mode.

[0030] If in work mode, activate the anti-deviation mode; in chat mode, engage in free conversation.

[0031] As a specific implementation of this application, if in working mode, the anti-deviation mechanism is activated, specifically including:

[0032] Target anchoring involves extracting core tasks as anchor points, storing them independently at set times, and synchronously saving snapshots of the technical environment.

[0033] Deviation monitoring: calculate the deviation value at each step of the discussion.

[0034] Tiered intervention is implemented, and the corresponding intervention plan is invoked based on the deviation value.

[0035] As a specific implementation of this application, when handling new topics, a deployed dialogue focus memory mechanism is also used. The specific operational logic includes:

[0036] Each time a user speaks, extract the core keywords and store them in the focus memory repository;

[0037] When the system generates a response, it performs a self-check to see if it contains keywords from the focus memory repository;

[0038] If it is included, then a normal response will be given;

[0039] If not included, the response content will be automatically corrected.

[0040] As a preferred implementation of this application, the progressive emotion mining tool is further used for:

[0041] When the matching of historical memory associations is insufficient to meet the preset threshold, the memory puzzle is completed through natural questioning to improve the accuracy of historical matching.

[0042] Secondly, embodiments of the present invention also provide a dialogue flow-based memory anchoring method, applied to a dialogue flow-based memory anchoring system as described in the first aspect, the method comprising the following steps:

[0043] Receive natural dialogue streams;

[0044] The natural dialogue flow is progressively analyzed to dynamically generate emotion tags;

[0045] The storage path is divided according to the emotion tag to realize the hierarchical storage memory on the end side; wherein, the emotion tree memory is constructed with tree storage nodes, and the physical isolation storage of multi-dimensional features is realized through the tree storage nodes;

[0046] Passive wake-up detection is performed based on the natural dialogue flow to distinguish the corresponding scenario and determine whether historical memory association is triggered; wherein, the scenario includes instant communication and historical memory.

[0047] If triggered, the system will intelligently retrieve historical memories and output a response.

[0048] If not triggered, a new entity is recorded and stored, and a corresponding natural response is generated.

[0049] The technical solution provided by the embodiments of the present invention has the following effects and advantages:

[0050] 1. Through progressive emotion mining and storage in an emotion tree structure, the tree structure enables cross-temporal and spatial association, achieving memory continuity and eliminating memory depth gaps; the emotion root node is permanently anchored in local flash memory, breaking through the traditional 128K token limit;

[0051] 2. By using passive wake-up based on the natural dialogue flow, historical memories are associated and integrated into the natural dialogue flow, thereby eliminating the feeling of mechanical interaction;

[0052] 3. Breaking through energy consumption limitations: Anchoring can be achieved through the end side, effectively reducing energy consumption. Attached Figure Description

[0053] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0054] Figure 1 This is a block diagram illustrating the principle of a dialogue-based memory anchoring system provided in an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the architecture of a memory anchoring system based on dialogue flow provided in an embodiment of the present invention;

[0056] Figure 3 This is a flowchart of a dialogue flow-based memory anchoring method provided in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0059] Please refer to Figure 1 This invention provides a dialogue-stream-based memory anchoring system applied to a smart terminal. The system includes:

[0060] A multimodal input interface for receiving natural dialogue streams;

[0061] A progressive sentiment miner is used to progressively analyze the natural dialogue stream and dynamically generate sentiment tags.

[0062] An emotion tree-structured memory is used to divide storage paths according to the emotion tags in order to achieve hierarchical storage memory on the end side; wherein, the emotion tree-structured memory is constructed with tree-structured storage nodes, and the physical isolation storage of multi-dimensional features is achieved through the tree-structured storage nodes;

[0063] A passive wake-up engine is used to perform passive wake-up detection based on the natural dialogue flow, so as to distinguish the corresponding scenario and thus determine whether the historical memory association is triggered; wherein, the scenario includes instant communication and historical memory.

[0064] A natural dialogue generator for:

[0065] If triggered, the system will intelligently retrieve historical memories and output a response.

[0066] If not triggered, a new entity is recorded and stored, and a corresponding natural response is generated.

[0067] During implementation, users input natural dialogue content into the system through voice, text, or images (such as speaking, typing, or sending pictures);

[0068] After receiving user input, the system passes the content to the "multimodal input interface." This interface performs preliminary processing on different forms of input (speech, text, and images), such as converting speech to text and recognizing key information in images, so that subsequent modules can better understand it.

[0069] The progressive analysis, which dynamically generates sentiment labels, specifically includes:

[0070] The process involves shallow feature extraction, mid-level semantic parsing, and deep sentiment fusion. A sentiment value is calculated based on a preset weight adaptive rule, and a corresponding sentiment label is generated based on the magnitude of the sentiment value. The sentiment label includes strong events and weak events.

[0071] 1. Voiceprint spectrum analysis: Fundamental frequency <180Hz → nostalgia, >220Hz → anger; i.e., shallow feature extraction;

[0072] 2. Text sentiment analysis: Keyword weighting algorithm (e.g., "happy" +0.3, "annoyed" +0.7); i.e., mid-level semantic analysis;

[0073] 3. Bimodal fusion formula: E = 0.7 × Evoice + 0.3 × Etext; that is, deep emotional fusion.

[0074] The weight adaptive rules include:

[0075] Plain text input → Text analysis weight 100%;

[0076] Pure voice input → Voiceprint analysis weight 100%;

[0077] Mixed input → 70% voiceprint + 30% text.

[0078] Storage paths are divided based on the emotion tags to achieve hierarchical memory storage on the client side, specifically including:

[0079] Strong events: Store the raw data in the "emotion root node to strong emotion area" according to the four-level structure of "time branch to scene leaf to entity leaf to feature node";

[0080] Weak events: compressed into 128-dimensional feature vectors and stored in the "emotion root node to weak emotion area to feature vector pool".

[0081] Specifically, storage paths are divided based on the emotion value E; E ≥ 0.8 can be interpreted as strong events, and weak events (E < 0.8);

[0082] That is, using emotion value as the basis for storage partitioning (strong / weak event areas), and using tree-like nodes to achieve physical isolation storage of four-dimensional features: time, space, entity, and emotion.

[0083] 1. Emotion root node partitioning: E≥0.9 (anger), 0.8≤E<0.9 (nostalgia), etc.;

[0084] 2. Dynamic attachment mechanism: New conversations are automatically associated with the closest emotion node;

[0085] 3. Storage optimization strategy: Store raw data for strong emotional events and feature vectors for weak events.

[0086] Strong event storage processing: Save raw data (photos / videos) and generate tag indexes (emotion / time / scene).

[0087] Weak event storage processing: Generate a 128-dimensional feature vector and perform BFP16 format compression.

[0088] Data is uniformly collected: the results of strong / weak event processing are ultimately stored in a tree-structured memory network.

[0089] Table 1 provides a detailed description of the storage.

[0090] hierarchy Node type Data types Example Storage strategy Level 0 node Emotional root node Partitioned Index Strong Emotion Zone / Weak Emotion Zone Memory resident Level 1 node Time branch node Timestamp 202x Flash encrypted storage Level 2 nodes Scene leaf node Scene tags graduation Flash Compressed Storage Level 3 nodes Entity leaf node Object Name red dress Scene-bound storage Level 4 nodes Feature nodes Detailed description Pearls on cuffs Dedicated storage for strong events - Feature vector pool 128-dimensional vector Histogram of main colors of roses Weak event pool

[0091] Table 2 Comparison of Storage Rules

[0092] characteristic Strong emotion zone storage Storage of weak emotion zones Data Depth Complete storage of level 4 nodes Single-layer feature pool storage Space usage Approximately 2.5KB / event 0.5KB / event Search speed Average 5ms Average 2ms Access frequency High frequency (≥ times / week) Low frequency (≤ times / month)

[0093] Meanwhile, an extreme compression storage strategy is adopted, as shown in Table 3.

[0094] Table 3 Compressed Storage Contents

[0095]

[0096] Furthermore, triggering historical memory associations includes:

[0097] Keyword trigger: Preset keywords are detected, and related events within ± preset years are retrieved based on the current sentiment value.

[0098] Silence trigger: When the current conversation is paused for more than a set time, a light scan of multiple recent topic keywords is performed; and when related, priority is given to matching those with high emotional similarity and core entities.

[0099] In this embodiment, historical memory association is triggered only in two cases:

[0100] Keyword trigger: When words such as "last time", "last year", and "do you still remember" are detected in the wakefulness word library, related events within ±(4-3E) years are retrieved based on the current sentiment value E (e.g., when E=0.85, ±1.45 years are retrieved).

[0101] Silence Trigger: When the conversation pauses for more than 10 seconds, perform a light scan of the 3 most recent topic keywords.

[0102] When making associations, priority is given to matching entities with high emotional similarity (S = 0.7E_now + 0.3E_hist) and core entities (such as personal names > common nouns).

[0103] If historical memory is triggered: retrieve details of the matching event and generate a natural response;

[0104] If it is a new topic: the system silently records the new entity, generates a brief contextual response, and does not force the association with history.

[0105] Process ended

[0106] After the conversation ends, all memories are automatically archived to the corresponding tree-structured storage node, waiting for the next conversation to be triggered, without any user intervention.

[0107] Reference Figure 2 Please provide an explanation of the text in the image:

[0108] ①Natural user dialogue:

[0109] Receiving everyday conversational phrases (such as an elderly person remarking, "I had so much fun the day my granddaughter danced").

[0110] Supports voice / text input, no command keywords required;

[0111] ② Real-time response generation:

[0112] Responses are not linked to historical memories: They are generated solely based on the current conversation content.

[0113] To prevent topics from veering off-topic: Users can freely switch topics;

[0114] Example response: "Yes, kids are so cute when they dance."

[0115] ③Memory storage:

[0116] Three key elements for automatic recording:

[0117] Emotional labels (e.g., "happy");

[0118] Entity tags (such as "granddaughter" or "dancing");

[0119] Time stamp (automatically adds a timestamp "202x-xx-xx");

[0120] It does not trigger any response; it is only stored in the background.

[0121] ④ Passive wake-up mechanism (dashed arrow):

[0122] Activate only when the user explicitly mentions history (e.g., "last time" or "before");

[0123] Wake-up example: The user says "I remember the dress I wore when I was dancing" → retrieve the stored "red dress" tag.

[0124] ⑤ Natural Language Output:

[0125] The response from the merged memory: "You mean the one that looks like a flower when it spins?"

[0126] Maintain your usual conversational style.

[0127] In summary, the system is described with reference to Table 4.

[0128] Table 4 System Structure and Function Description

[0129]

[0130] It should be noted that strong events are stored in their original details (four-level tree structure), while weak events are stored in feature vectors (pooled storage).

[0131] Triggering principle: Only respond passively to explicit historical indications (keywords / silence) from users, without actively asking follow-up questions;

[0132] Freedom of conversation: Users can switch topics at will, and the system only silently records new information to ensure that the interaction is as natural as casual conversation.

[0133] The entire process, from user input to memory archiving, is completed entirely on the device without cloud support. The core is to achieve the integration of long-term memory and natural dialogue through "emotional tags + tree-structured storage + passive wake-up".

[0134] Furthermore, based on the above technical solution, in another embodiment, the system further includes an anti-deviation module, which is used for:

[0135] If historical memory is triggered, the original target anchor point will be automatically activated, anti-deviation monitoring will be started, and a natural response will be generated.

[0136] If it is a new topic for instant communication, the mode is determined first; the mode includes work mode and casual chat mode.

[0137] If in work mode, activate the anti-deviation mode; in chat mode, engage in free conversation.

[0138] During implementation, if in working mode, activate anti-deviation measures, which specifically include:

[0139] Target anchoring involves extracting core tasks as anchor points, storing them independently at set times, and synchronously saving snapshots of the technical environment.

[0140] Deviation monitoring: calculate the deviation value at each step of the discussion.

[0141] Tiered intervention is implemented, and the corresponding intervention plan is invoked based on the deviation value.

[0142] The working mode (anti-deviation startup) is as follows: when the following conditions are met: "≥1 professional keyword detected (such as "implementation" or "error reporting") + verb extracted + noun combination (such as "fix login bug") + context contains technical description";

[0143] Target anchoring: Extract core tasks as anchor points (such as "file upload repair"), store them independently for 72 hours, and simultaneously save a snapshot of the technical environment;

[0144] Deviation monitoring: Deviation value is calculated at each discussion step = (Repair step depth / 5) × (1 - Keyword matching degree);

[0145] Tiered intervention:

[0146] Slight deviation (30%-50%): The prompt reads, "The current problem does not directly affect the [core objective]. Should we continue?".

[0147] Moderate deviation (50%-70%): Warning " "May delay [core objective] by more than 2 hours";

[0148] Severe deviation (>70%): Provide three options (revert to basic version / alternative version / continue the current path).

[0149] Casual Chat Mode (No Intervention): When the following conditions are met: "3 consecutive sentences without professional terms + containing everyday vocabulary (such as weather, food) + emotion value E < 0.4", the system will only silently record new entities (such as "fog" or "picking flower petals"), generate a short contextual response (such as "kids love to play like this"), and does not restrict topic switching.

[0150] Process ended

[0151] After the conversation ends, all memories are automatically archived to the corresponding tree-structured storage node (strong event four-level structure / weak event feature pool), and the anti-deviation anchor point is archived synchronously, waiting for the next conversation to be triggered, without any user intervention.

[0152] Summary of core logic

[0153] Storage objective: Strong events are stored in their original details (four-level tree structure), and weak events are stored in their feature vectors (pooled storage);

[0154] Triggering principle: Only respond passively to explicit historical indications (keywords / silence) from users, without actively asking follow-up questions;

[0155] Mode adaptation: Work mode starts to prevent deviation (anchoring the target + tiered intervention), chat mode allows for free conversation (zero intervention);

[0156] Freedom of conversation: Users can switch topics at will, and the system only silently records new information to ensure that the interaction is as natural as casual conversation.

[0157] The entire process, from user input to memory archiving, is completed entirely on the device without cloud support. The core is to achieve long-term memory and natural dialogue through "emotional tags + tree-structured storage + passive wake-up + pattern adaptation".

[0158] Furthermore, when handling new topics, a dialogue focus memory mechanism is deployed, the specific operational logic of which includes:

[0159] Each time a user posts, the core keywords are extracted and stored in the focus memory repository (capacity: the 3 most recent keywords);

[0160] When the system generates a response, it performs a self-check to see if it contains keywords from the focus memory repository;

[0161] If it is included, then a normal response will be given;

[0162] If not included, the response content will be automatically corrected.

[0163] The self-correction process includes:

[0164] U [User Comments] --> S [System]

[0165] S --> | Current Round | F [Focus Memory Container]

[0166] F --> |Storage|K [Keyword "flower"]

[0167] S-->|Next Round Response|C{Focus Matching}

[0168] C --> | Match successful | R [Continuing the topic]

[0169] C --> | Match failed | M [Self-correcting]

[0170] Focus memory capacity: only the 3 most recent keywords are saved (memory <0.1KB); Scan frequency: only scan once before the response is generated (no continuous monitoring); Correction time: <5ms (imperceptible).

[0171] For ease of understanding, the following is an explanation of the natural dialogue memory anchoring interaction process: First round: Trigger word response User input: "When I saw the flowers last year..." (including the historical trigger word "last year") System response:

[0172] The engine is activated to retrieve matching events → the memory for "202x Park Roses" is returned, generating an associated response: "Did Niu Niu pull the pink roses?"

[0173] Round Two: Silence Response

[0174] User behavior: Silence for 11 seconds

[0175] System response:

[0176] Activate lightweight topic anchoring scan

[0177] The risk of topic deviation is judged to be low (<30%).

[0178] The topic naturally continued: "Later, the petals..."

[0179] Round 3: New Topic Handling User Input: "Is the child naughty?" (Switch to a new topic)

[0180] System response:

[0181] Record the new entity "child" to memory.

[0182] Passive awakening of associated memories → Generation of a response: "You always say she's making a fuss."

[0183] Furthermore, the memory association and response process in natural dialogue.

[0184] First round of dialogue: Real-time recording and response

[0185] User → System input: "The flowers in the park are so beautiful" (natural expression, pleasant tone).

[0186] System → Memory Storage

[0187] Record 3D labels:

[0188] 1. Emotion: Pleasant (E=0.7);

[0189] 2. Entity: Flower;

[0190] 3. Time: 202x-xx-xx (current date).

[0191] 3. System → User

[0192] Response: "Seeing flowers in spring makes me feel good" (an immediate empathetic response based on the current context).

[0193] Second round of dialogue: Awakening historical memory

[0194] User → System

[0195] Input: "I brought Niu Niu here last year" (actively mentioning a past event).

[0196] System → Memory Storage

[0197] Keyword detection: Extracting "last year" triggers time-based search, and "Niu Niu" triggers entity-based search.

[0198] Memory storage → system

[0199] Return to history: Memories related to "Niu Niu + Park" in 202x-xx-month (may include details such as "picking flowers").

[0200] System → User Response: "Did your granddaughter pull at the rose bush?" (Linking historical memories and accurately recalling details).

[0201] Third round of dialogue: New information logging and lightweight response

[0202] User → System

[0203] Enter: "She always loves to pick flowers" (start a new topic and add the entity behavior).

[0204] System → Memory Storage

[0205] New tag: Entity "Picking Flowers", linked to "Niu Niu" historical archive.

[0206] System → User Response: "All children are like this" (Does not force association with history, adopts a universal empathetic response).

[0207] In summary, its core processing logic is shown in Table 5.

[0208] Table 5 Summary of Core Logic

[0209]

[0210]

[0211] To more intuitively illustrate the differences between this solution and the traditional solution, please refer to Table 6.

[0212] Table 6. Technology Comparison Table

[0213]

[0214] Furthermore, in another embodiment, based on the foregoing scheme, the progressive emotion mining tool is also used for:

[0215] When the matching of historical memory associations is insufficient to meet the preset threshold, the memory puzzle is completed through natural questioning to improve the accuracy of historical matching.

[0216] That is, the system is triggered when a user recalls history but the description is too vague; the system will naturally ask for details like chatting with a friend to complete the memory puzzle and improve the accuracy of history matching; see Table 7.

[0217] For example, when the matching is insufficient to meet the threshold (similarity < 0.65), features are dynamically supplemented through three levels of natural follow-up questions: entity → spatiotemporal → emotion, thereby improving the hit rate of weak event retrieval.

[0218] Primarily used to initiate multi-level emotion-guided follow-up questions when historical dialogue matching fails. It improves feature matching accuracy by dynamically generating incremental keywords, thereby increasing the accuracy of historical dialogue matching. Specifically, it includes:

[0219] The follow-up questioning process is triggered when the similarity of the vector pool retrieval is <0.65.

[0220] Follow-up questions are generated in a three-level order: basic entity → spatiotemporal attributes → emotional characteristics;

[0221] (Example: Level 1 "What type of flower are you talking about?" → Level 2 "In which month did you see it?" → Level 3 "You seemed quite happy at the time, didn't you?")

[0222] Table 7 Intelligent Questioning Strategies

[0223] User omissions System follow-up questioning method Friend chat analogy He didn't say what it was about. Are you referring to a meal or a gathering? Is it a dinner date or a birthday party? No location mentioned "In the city center or the west side?" "Is it the one by West Lake?" No details were given. Were you happy at the time? "Having fun, huh?"

[0224] The above solution has the following effects and advantages:

[0225] 1. Through progressive emotion mining and storage in an emotion tree structure, the tree structure enables cross-temporal and spatial association, achieving memory continuity and eliminating memory depth gaps; the emotion root node is permanently anchored in local flash memory, breaking through the traditional 128K token limit;

[0226] 2. By using passive wake-up based on the natural dialogue flow, historical memories are associated and integrated into the natural dialogue flow, thereby eliminating the feeling of mechanical interaction;

[0227] 3. Breaking through energy consumption limitations: Anchoring can be achieved through the end side, effectively reducing energy consumption.

[0228] Based on the same inventive concept, and referring to Figure 3 The present invention also provides a dialogue flow-based memory anchoring method, applied to a dialogue flow-based memory anchoring system as described in the first aspect, the method comprising the following steps:

[0229] S101, receives natural dialogue stream;

[0230] S102, perform progressive analysis on the natural dialogue flow and dynamically generate emotion tags;

[0231] S103, divide the storage path according to the emotion tag to realize the hierarchical storage memory on the terminal side; wherein, the emotion tree memory is constructed with tree storage nodes, and the physical isolation storage of multi-dimensional features is realized through the tree storage nodes;

[0232] S104, Passive wake-up detection is performed based on the natural dialogue flow to distinguish the corresponding scenario, thereby determining whether historical memory association is triggered; wherein, the scenario includes instant communication and historical memory;

[0233] S105, if triggered, will perform intelligent recall of historical memory association and output a response;

[0234] S106, if not triggered, record and store the new entity and generate the corresponding natural response.

[0235] Furthermore, the progressive analysis, which dynamically generates sentiment labels, specifically includes:

[0236] The process involves shallow feature extraction, mid-level semantic parsing, and deep sentiment fusion. A sentiment value is calculated based on a preset weight adaptive rule, and a corresponding sentiment label is generated based on the magnitude of the sentiment value. The sentiment label includes strong events and weak events.

[0237] Storage paths are divided based on the emotion tags to achieve hierarchical memory storage on the client side, specifically including:

[0238] Strong events: Store the raw data in the "emotion root node to strong emotion area" according to the four-level structure of "time branch to scene leaf to entity leaf to feature node";

[0239] Weak events: compressed into 128-dimensional feature vectors and stored in the "emotion root node to weak emotion area to feature vector pool".

[0240] Triggering historical memory associations includes:

[0241] Keyword trigger: Preset keywords are detected, and related events within ± preset years are retrieved based on the current sentiment value.

[0242] Silence trigger: When the current conversation is paused for more than a set time, a light scan of multiple recent topic keywords is performed; and when related, priority is given to matching those with high emotional similarity and core entities.

[0243] The method further includes:

[0244] If historical memory is triggered, the original target anchor point will be automatically activated, anti-deviation monitoring will be started, and a natural response will be generated.

[0245] If it is a new topic for instant communication, the mode is determined first; the mode includes work mode and casual chat mode.

[0246] If in work mode, activate the anti-deviation mode; in chat mode, engage in free conversation.

[0247] Furthermore, the method also includes:

[0248] When the matching of historical memory associations is insufficient to meet the preset threshold, the memory puzzle is completed through natural questioning to improve the accuracy of historical matching.

[0249] That is, when a user recalls history but the description is too vague, it is triggered; the system will naturally ask for details like chatting with a friend, to complete the memory puzzle and improve the accuracy of history matching;

[0250] For example, when the matching is insufficient to meet the threshold (similarity < 0.65), features are dynamically supplemented through three levels of natural follow-up questions: entity → spatiotemporal → emotion, thereby improving the hit rate of weak event retrieval.

[0251] It should be noted that for a more detailed description of the workflow of the method embodiments, please refer to the aforementioned system embodiments section, which will not be repeated here.

[0252] The entire solution achieved the following:

[0253] ① Progressive emotion mining achieves a qualitative leap in emotion understanding through three-level analysis of voiceprint and text, breaking through the traditional single feature recognition;

[0254] ② Emotion tree-like storage, constructing a four-level storage structure to achieve permanent memory anchoring and eradicate the problem of dialogue gaps;

[0255] ③ Passive wake-up mechanism, based on trigger word library to ensure natural dialogue flow.

[0256] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A memory anchoring system based on dialogue flow, characterized in that, The system, applied to smart terminals, includes: A multimodal input interface for receiving natural dialogue streams; A progressive sentiment miner is used to progressively analyze the natural dialogue stream and dynamically generate sentiment tags. An emotion tree-structured memory is used to divide storage paths according to the emotion tags in order to achieve hierarchical storage memory on the end side; wherein, the emotion tree-structured memory is constructed with tree-structured storage nodes, and the physical isolation storage of multi-dimensional features is achieved through the tree-structured storage nodes; A passive wake-up engine is used to perform passive wake-up detection based on the natural dialogue flow to distinguish the corresponding scenario and thus determine whether to trigger historical memory association; wherein, the scenario includes instant communication and historical memory. A natural dialogue generator for: If triggered, the system will intelligently retrieve historical memories and output a response. If not triggered, a new entity is recorded and stored, and a corresponding natural response is generated.

2. The memory anchoring system based on dialogue flow as described in claim 1, characterized in that, The progressive analysis, which dynamically generates sentiment labels, specifically includes: The process involves shallow feature extraction, mid-level semantic parsing, and deep sentiment fusion. A sentiment value is calculated based on a preset weight adaptive rule, and a corresponding sentiment label is generated based on the magnitude of the sentiment value. The sentiment label includes strong events and weak events.

3. A memory anchoring system based on dialogue flow as described in claim 1, characterized in that, Storage paths are divided based on the emotion tags to achieve hierarchical memory storage on the client side, specifically including: Strong events: Store the raw data in the "emotion root node to strong emotion area" structure, and store the raw data in a four-level structure of "time branch to scene leaf to entity leaf to feature node". Weak events: compressed into 128-dimensional feature vectors and stored in the "emotion root node to weak emotion area to feature vector pool".

4. A memory anchoring system based on dialogue flow as described in claim 3, characterized in that, Triggering historical memory associations includes: Keyword trigger: Preset keywords are detected, and related events within ± preset years are retrieved based on the current sentiment value. Silence trigger: When the current conversation is paused for more than a set time, a light scan of multiple recent topic keywords is performed; and when related, priority is given to matching those with high emotional similarity and core entities.

5. A dialogue-flow-based memory anchoring system as described in any one of claims 1 to 4, characterized in that, The system also includes an anti-deviation module, which is used for: If historical memory is triggered, the original target anchor point will be automatically activated, anti-deviation monitoring will be started, and a natural response will be generated. If it is a new topic for instant communication, the mode is determined first; the mode includes work mode and casual chat mode. If in work mode, activate the anti-deviation mode; in chat mode, engage in free conversation.

6. A dialogue-flow-based memory anchoring system as described in claim 5, characterized in that, If in working mode, activate anti-deviation measures, specifically including: Target anchoring involves extracting core tasks as anchor points, storing them independently at set times, and synchronously saving snapshots of the technical environment. Deviation monitoring: calculate the deviation value at each step of the discussion. Tiered intervention is implemented, and the corresponding intervention plan is invoked based on the deviation value.

7. A dialogue-flow-based memory anchoring system as described in claim 5, characterized in that, When handling new topics, a dialogue focus memory mechanism is also deployed. The specific operational logic includes: Each time a user speaks, extract the core keywords and store them in the focus memory repository; When the system generates a response, it performs a self-check to see if it contains keywords from the focus memory repository; If it is included, then a normal response will be given; If not included, the response content will be automatically corrected.

8. A dialogue-flow-based memory anchoring system as described in claim 5, characterized in that, The progressive emotion mining tool is also used for: When the matching of historical memory associations is insufficient to meet the preset threshold, the memory puzzle is completed through natural questioning to improve the accuracy of historical matching.

9. A memory anchoring method based on dialogue flow, characterized in that, The method applied to the dialogue-flow-based memory anchoring system of claim 1 includes the following steps: Receive natural dialogue streams; The natural dialogue flow is progressively analyzed to dynamically generate emotion tags; The storage path is divided according to the emotion tag to realize the hierarchical storage memory on the end side; wherein, the emotion tree memory is constructed with tree storage nodes, and the physical isolation storage of multi-dimensional features is realized through the tree storage nodes; Passive wake-up detection is performed based on the natural dialogue flow to distinguish the corresponding scenario and determine whether historical memory association is triggered; wherein, the scenario includes instant communication and historical memory. If triggered, the system will intelligently retrieve historical memories and output a response. If not triggered, a new entity is recorded and stored, and a corresponding natural response is generated.

10. A memory anchoring method based on dialogue flow as described in claim 9, characterized in that, The method further includes: If historical memory is triggered, the original target anchor point will be automatically activated, anti-deviation monitoring will be started, and a natural response will be generated. If it is a new topic for instant communication, the mode is determined first; the mode includes work mode and casual chat mode. If in work mode, activate the anti-deviation mode; in chat mode, engage in free conversation.

Images