An old-age care intelligent agent personalized intervention system based on long-term memory and closed-loop feedback self-evolution

Abstract

The application discloses a kind of based on long-term memory and closed loop feedback self-evolution's old-age care intelligent agent personalized intervention system.It relates to the technical field of intelligent old-age care.The system includes: the multiple-source event access module connected in turn, structured event analysis module, long-term memory recall module, memory credibility arbitration module, intervention ticket generation and issue module, closed loop feedback module, self-evolution update module and long-term memory engine, structured event analysis module, long-term memory recall module are also connected with long-term memory engine.The application can significantly reduce the cost of artificial maintenance, improve the continuity, accuracy, stability, continuous adaptation ability and strategy evolution ability of the system in the long-term operation of old-age care task.

Description

Technical Field

[0001] This invention relates to the field of intelligent elderly care technology, and in particular to a personalized intervention system for elderly care intelligent agents based on long-term memory and closed-loop feedback self-evolution. Background Technology

[0002] Currently, the demand for continuous companionship, health reminders, emotional support, attention to abnormalities, and individualized intervention is rapidly increasing in home-based, community-based, and institutional elderly care settings. Existing elderly care companionship systems typically provide services through voice interaction, fixed rule reminders, schedule management, or simple user profiling mechanisms. While these systems can handle basic question-and-answer, broadcasting, and reminders, they still have significant shortcomings in providing long-term, continuous service.

[0003] In existing technologies, while some solutions establish user profiles, these are mostly static data, primarily used for display or simple matching, and are difficult to play a dynamic role in specific intervention decisions. In other words, even if the system "remembers the user," it does not truly "use memory to form more appropriate care interventions."

[0004] Existing systems' "adaptive" capabilities often remain at the level of manually modifying prompts, adjusting a few rules, or manually maintaining user profiles. They lack a self-evolving mechanism that is based on long-term memory, relies on closed-loop feedback, and automatically updates memory content and strategy parameters. This makes it difficult to gradually adapt to the specific needs and behavioral characteristics of elderly users during long-term care.

[0005] Therefore, there is a need to provide a new personalized intervention method for elderly care companionship to improve the continuity, accuracy, individual adaptability, and long-term optimization capabilities of elderly care companionship tasks. Summary of the Invention

[0006] The purpose of this invention is to provide a personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, aiming to solve the aforementioned problems and improve the continuity, accuracy, individual adaptability, and long-term optimization capabilities of elderly care companionship tasks. The various technical effects of the preferred technical solutions provided by this invention are detailed below.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, comprising: The multi-source event access module encapsulates the received multi-source care data into standard care event objects; The structured event parsing module performs event classification, risk grading, emotion discrimination, time binning, and scene labeling on the standard caregiver event objects to obtain the parsing results; The long-term memory retrieval module performs structured index retrieval, semantic representation retrieval, and relation expansion retrieval in the long-term memory engine based on the structure fields corresponding to the parsing results, the summary of the standard care event object, the key entity tags extracted from the standard care event object, historical intervention tickets, and known event relationships maintained by the long-term memory engine. The retrieval results are then fused and rearranged to obtain the activated memory item of the standard care event object. The memory credibility arbitration module calculates the comprehensive credibility score of the activated memory item based on the source weight, time validity, verification consistency, and conflict penalty item, and removes memory items that are below the threshold. The intervention ticket generation and distribution module generates a personalized intervention plan consisting of multiple intervention tickets based on the effect memory stored in the long-term memory engine for the active memory items that have not been removed, and selects the intervention ticket with the highest priority to distribute and execute. The closed-loop feedback module forms an execution evidence package from the continuously received feedback information after the intervention ticket is executed, and writes the execution evidence package back to the long-term memory engine and the strategy version record table respectively. The self-evolution update module triggers event-level real-time updates and periodic-level batch updates based on the execution evidence package.

[0008] In some embodiments, for the newly added care information corresponding to the standard care event object, the long-term memory engine performs field standardization, key entity extraction, source marking, conflict detection and credibility initialization, and then generates the corresponding semantic representation, which is written into the relational table, semantic index and relation association index respectively.

[0009] In some embodiments, the semantic representation retrieval includes: The summary text of each memory item is encoded as a dense vector representation of a fixed dimension, and the vector is written into a vector index library; the summary of the standard care event object is encoded as a dense vector, and the most semantically relevant memory item is recalled from the vector index library as the retrieval result using cosine similarity as a metric.

[0010] In some embodiments, the search results are merged and rearranged, including: The comprehensive score of each historical memory item in the search results is calculated based on the structural matching score, semantic similarity score, relational association score, credibility score, and time decay score corresponding to the historical memory item; based on the calculated comprehensive score of the historical memory item, historical memory items with a comprehensive score higher than a preset threshold are selected to constitute the activated memory items of the standard care event object; The structure matching score is determined based on the matching degree of event type and scene tag of each memory item in the search results. The semantic similarity score is determined based on the vector cosine similarity between the summary of the standard caregiver event object and the summary of each memory item in the search results. The relational association score is determined based on the overlap between the key entity tags involved in the standard caregiver event object and the key entity tags associated with each memory item in the search results. The credibility score is taken from the comprehensive credibility score of each memory item in the search results obtained through the memory credibility arbitration module. The timeliness decay score is calculated based on the interval between the last update time and the current time of each memory item in the search results using an exponential decay function.

[0011] In some embodiments, the long-term memory engine includes tables that hierarchically store fact memory, preference memory, intervention memory, and effect memory; wherein, the fields of the fact memory table include user identifier, fact key name, fact value, source, credibility, and update time; the fields of the preference memory table include preference type, preference style, applicable time range, and credibility; the fields of the intervention memory table include trigger type, action type, result label, and failure reason; and the fields of the effect memory table include action type, scene label, time bucket, acceptance rate, aversion score, decay factor, and recovery period. The long-term memory engine also includes a hot memory layer, a steady-state memory layer, and an archive memory layer; wherein, the hot memory layer is used to store recently frequently called and currently effective memory items, the steady-state memory layer is used to store currently effective fact memory, preference memory, intervention memory, and effect memory, and the archive memory layer is used to store low-frequency historical events, expired memories, and versioned execution evidence, and performs segmented archiving according to time windows.

[0012] In some embodiments, a memory maintenance module is further included for maintaining the memory data of the long-term memory engine, wherein the memory maintenance includes: Traverse the memory items in the steady-state memory layer that have exceeded the preset validity period and have not been recently verified in closed loop, and update their status according to the verification results; Identify multiple memory items with different sources and contradictory content under the same user and the same fact key, and determine the retention or verification status based on priority scoring; For multiple memory items whose cosine similarity between summary vectors exceeds a preset merging threshold, have the same source, and have consistent closed-loop verification results, merge them while retaining the latest timestamp and the highest credibility. The steady-state memory item copy whose recent call frequency exceeds the hotness threshold is promoted to the hot memory layer cache. The memory items in the hot memory layer that have not been called for multiple consecutive cycles are removed from the cache and only the steady-state layer copy is retained.

[0013] In some embodiments, the long-term memory engine updates the parameters of the effect memory of the activated memory item corresponding to the intervention ticket based on the feedback information continuously received by the closed-loop feedback module after the intervention ticket is executed; the parameters of the effect memory include initial acceptance, recent acceptance, decay parameter, aversion threshold, recovery period and scene adaptation weight. The update uses a combination of incremental statistics and time decay, including: First, the activated memory items are statistically analyzed by time period, scene tag, and emotion tag. Then, the acceptance rate, completion rate, and aversion score of the activated memory items are updated exponentially.

[0014] In some embodiments, when the overall credibility of the activated memory item is lower than the execution threshold, and its risk level belongs to a high-risk scenario or a conflict scenario, the action-type intervention to be executed is automatically switched to a confirmation-type intervention, and a secondary confirmation request is initiated to the user, family member, or caregiver.

[0015] In some embodiments, the intervention ticket includes a standard caregiver event object identifier, the key memory item invoked, the recommended intervention method, the expected execution timing, the expected feedback indicators, the recurrence triggering conditions, the escalation threshold, and the review requirements; The execution evidence package includes the corresponding intervention ticket identifier, execution time, user's immediate response, whether it was accepted, whether it was completed, completion delay, whether a reminder is needed, whether a negative reaction occurred, whether it entered escalation processing, manual confirmation result, and final effect label; The feedback information includes user language feedback, changes in environmental conditions, records of care task completion, family member confirmation, caregiver confirmation, and / or periodic review results.

[0016] In some embodiments, the self-evolutionary update module includes a micro-evolutionary layer and a macro-evolutionary layer; the micro-evolutionary layer is used to locally update the memory items involved and their effect memory parameters after a single intervention event ends; the macro-evolutionary layer is used to optimize the system-level parameters as a whole after multiple rounds of intervention events or a preset period, the system-level parameters including intervention selection weights, risk thresholds under different scenarios, repeated reminder limits, upgrade triggering conditions, memory expiration period, and effect decay model parameters; The self-evolutionary update module adopts a memory lifecycle management mechanism. When a memory item has not been verified for a long time, has repeatedly conflicted with closed-loop feedback, or is no longer applicable in the current stage, the self-evolutionary update module automatically reduces its effective weight or archives it. When the effect of a certain type of intervention scheme continues to deteriorate, the self-evolutionary update module automatically reduces its recommended scope or marks it as a high-disruption method. The memory lifecycle includes six states: generation, activation, verification, demotion, freezing, and archiving.

[0017] Implementing one of the above-described technical solutions of the present invention has the following advantages or beneficial effects: This invention, by setting up intervention tickets, effect memory, and execution evidence packages, ensures that each personalized intervention has a complete chain of "traceable generation basis, recordable execution process, and rewritable execution results," thereby significantly enhancing the verifiability, auditability, and targeted optimization of elderly care tasks. This invention can significantly reduce manual maintenance costs and improve the system's continuity, accuracy, stability, continuous adaptability, and strategy evolution capabilities in the long-term operation of elderly care tasks. Attached Figure Description

[0018] The accompanying drawings used are briefly described below. It is obvious that the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 This is a block diagram of a personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, according to an embodiment of the present invention. Figure 2 This is a long-term memory engine framework diagram according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments, illustrating various exemplary embodiments that may be used to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of the present invention disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of the present invention.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "multiple" means two or more. The terms "connected" and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, integral connections, mechanical connections, electrical connections, communication connections, direct connections, indirect connections through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] To illustrate the technical solution described in this invention, specific embodiments are described below, showing only the parts related to the embodiments of this invention.

[0022] Example 1: like Figure 1 , Figure 2 As shown, this invention provides a personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution. This system can be applied to elderly care companionship robots, home-based elderly care terminals, home care platforms, community elderly care gateways, or other elderly care companionship systems. It establishes a personalized intervention architecture centered on long-term memory, updated based on closed-loop feedback, and continuously optimized through self-evolution. The system includes: a multi-source event access module, a structured event parsing module, a long-term memory retrieval module, a memory credibility arbitration module, an intervention ticket generation and distribution module, a closed-loop feedback module, a self-evolution update module, and a long-term memory engine, all connected sequentially. The structured event parsing module and the long-term memory retrieval module are also connected to the long-term memory engine. Furthermore, The multi-source event access module encapsulates the received multi-source caregiver data into standard caregiver event objects (hereinafter referred to as the current event or the corresponding event written into the long-term engine). These standard caregiver event objects can adopt a unified JSON structure or a relational record format, with fields including event identifier, user identifier, source type, collection time, original content, structured summary (generated from the original content), event type, scene tag, risk level, emotion tag, key entity tag, and evidence citation. For example, when the voice terminal receives the input "I don't want to take my blood pressure medication today," the system can generate a standard caregiver event object containing the user identifier, source type, event type, risk level, emotion tag, and key entity tag. This object is used to uniformly convert voice, text, sensor, and family-side messages into input that can be processed by the subsequent parsing, recall, and intervention modules.

[0023] The aforementioned evidence citations refer to the original or processed evidence identifiers that support the formation of the current standard caregiver event object, such as voice segment numbers, text message numbers, sensor record numbers, family member confirmation record numbers, caregiver record numbers, etc., which are used for subsequent auditing, retrospection, and credibility calculation.

[0024] In a specific embodiment, the sources of multi-source care data include care task records, family member supplementary records, caregiver supplementary records, abnormal event messages, environmental status change data, sensor events, and remote business messages, and their forms include user voice and text input through the interface.

[0025] The structured event parsing module performs event classification, risk grading, emotion identification, time binning, and scene labeling on standard caregiving events to obtain the parsing results. Event classification can include categories such as medication reminders, food and water intake, exercise rehabilitation, sleep schedules, emotional support, abnormal alarms, and family communication; risk grading can include low risk, medium risk, medium-high risk, and high risk; emotion categories can include stable, anxious, resistant, depressed, irritable, and seeking help; time binning can be divided into morning, forenoon, noon, afternoon, evening, night, or hourly windows; scene labeling can include scenarios such as home, institution, medical treatment, rehabilitation training, being alone, and family members online. In practice, the system can determine the event type and risk level based on keywords, source type, historical risk records, and sensor anomaly intensity, and output emotion and scene labels based on multi-source caregiving data using a semantic model or emotion dictionary.

[0026] The aforementioned sensor anomaly intensity indicates the degree of deviation of the sensor's current value from the user's baseline or a preset safety threshold. For example, a standardized anomaly score can be calculated using data such as heart rate, blood pressure, gait, falls, sleep, and activity levels, and this anomaly score can be used to identify event types and determine risk levels.

[0027] In a specific embodiment, risk classification can be divided into low risk, medium risk, medium-high risk, and high risk based on event risk scores. Event risk scores can range from 0 to 100 points and are obtained by weighting an event type base score, historical risk correction score, source reliability correction score, and sensor anomaly correction score. For example: The event risk score is calculated as follows: Event Type Base Score × 0.4 + Historical Risk Correction Score × 0.2 + Source Reliability Correction Score × 0.2 + Sensor Anomaly Correction Score × 0.2. The Event Type Base Score can be preset based on the event category, for example, 20-40 points for emotional support, 30-50 points for food and water, 40-60 points for exercise rehabilitation, 60-80 points for medication reminders, and 80-100 points for abnormal alarms. The Historical Risk Correction Score is determined based on the number of anomalies in similar events within the past 30 days, for example, 0 anomalies = 0 points, 1-2 anomalies = 40 points, and more than 3 anomalies = 80 points. The Source Reliability Correction Score is determined based on the source type, for example, 70 points for device sensing results, 80 points for caregiver records, 85 points for user input, and 90 points for family confirmation. The Sensor Anomaly Correction Score is determined based on the degree to which sensor data deviates from the user's historical baseline or safety threshold, for example, 0 points for no anomalies, 40 points for mild anomalies, 70 points for moderate anomalies, and 100 points for severe anomalies.

[0028] The risk level is further divided according to the risk score of the event: 0-30 is low risk, 31-60 is medium risk, 61-80 is medium-high risk, and 81-100 is high risk.

[0029] The long-term memory retrieval module performs structured index retrieval, semantic representation retrieval, and relation expansion retrieval in the long-term memory engine based on the structured fields corresponding to the parsing results, the summary of the standard caregiver event object, the key entity tags extracted from the standard caregiver event object, historical intervention tickets, and known event relationships maintained by the long-term memory engine. The retrieval results are then fused and rearranged to obtain the activated memory items of the standard caregiver event object.

[0030] In some embodiments, structured index retrieval precisely filters and retrieves historical memory items corresponding to standard caregiver event objects from the long-term memory engine based on the structure fields corresponding to the parsing results. Structure fields include user identifier, event type, scene tag, and time bucket.

[0031] Semantic representation retrieval expands candidates based on the semantic similarity between the summary of the standard caregiver event and the memory summary (the summary corresponding to the memory item stored in the long-term memory engine). Further, semantic representation retrieval includes: The summary text of each memory item is encoded as a dense vector representation of fixed dimensions, and the vector is written into a vector index library. The summary of standard caregiver event objects is encoded as a dense vector, and the most semantically relevant historical memory items are retrieved from the vector index library as search results using cosine similarity as a metric. The vector index library preferably employs an index structure based on approximate nearest neighbor search, such as an HNSW (Hierarchical Navigable Small World) index or an IVF (Inverted File) index, to achieve sublinear time complexity semantic retrieval in a large-scale set of memory items.

[0032] To put it simply, through the vectorized semantic representation retrieval mechanism, the system is able to recall semantically relevant historical memories even when keywords do not match perfectly.

[0033] Relationship expansion retrieval performs association expansion based on key entity tags, historical intervention tickets, and known event relationships related to the standard caregiver event object in the long-term memory engine, obtaining associated historical memory items for the standard caregiver event object. Key entity tags are identifiers extracted from the current event (standard caregiver event object), such as medications, care tasks, family members, diseases, time preferences, locations, or equipment; historical intervention tickets are intervention records generated by the system for similar events in the past, recording the intervention method, timing, and feedback evidence (see below); known event relationships are the associations between key entities, events, and intervention results maintained in the long-term memory engine. Relationship expansion retrieval can perform one-hop or multi-hop expansion based on the above information. For example, when the current event involves "blood pressure medication" and "evening," the system can retrieve past evening medication reminders, user resistance feedback, and family confirmation records, thereby recalling historical memory items related to the current event.

[0034] The aforementioned care tasks are health care and daily living care services provided by caregivers (including family members) to the elderly in places such as homes, hospitals, communities, and elderly care institutions. These services, such as medical and health care, can be determined in accordance with relevant laws, regulations, or actual circumstances.

[0035] The aforementioned intervention methods can indicate the degree of deviation of the sensor's current value from the user's baseline or a preset safety threshold. For example, a standardized anomaly score can be calculated using data such as heart rate, blood pressure, gait, falls, sleep, and activity levels, and this anomaly score can be used to identify event types and determine risk levels.

[0036] To put it simply, through the hybrid retrieval technology described above, the long-term memory engine can simultaneously achieve precise retrieval capabilities, semantic recall capabilities across expression forms, and historical association tracing capabilities.

[0037] In some embodiments, the fusion and rearrangement stage employs a weighted linear fusion strategy to comprehensively score the candidate memory items obtained from the hybrid retrieval. Specifically, The search results are merged and rearranged, including: The comprehensive score of each historical memory item in the search results is calculated based on the structural matching score, semantic similarity score, relational association score, credibility score, and time decay score corresponding to the historical memory items in the long-term memory engine. Based on the calculated comprehensive score, the historical memory items are sorted in descending order, and those with scores higher than a preset threshold are selected to constitute the activated memory items of the standard caregiver event object. Specifically, the structural matching score is determined by the degree of matching between the event type and scene tags of each memory item in the search results; the semantic similarity score is determined by the vector cosine similarity between the summary of the standard caregiver event object and the summary of each memory item in the search results; the relational association score is determined by the overlap between the key entity tags involved in the standard caregiver event object and the key entity tags associated with each memory item in the search results; the credibility score is taken from the comprehensive credibility score of each memory item in the search results obtained through the memory credibility arbitration module; and the time decay score is calculated using an exponential decay function based on the interval between the last update time and the current time of each memory item in the search results.

[0038] In a specific embodiment, a comprehensive scoring formula as follows: ; in, For structural matching, For semantic similarity, For relational classification, As a credibility score, For time-related decay, , , , , These are the weights for structural matching score, semantic similarity score, relational association score, credibility score, and timeliness decay score, respectively. These weights can be set according to the actual situation.

[0039] In some embodiments, for newly added caregiving information corresponding to standard caregiving event objects, a long-term memory engine performs field standardization, key entity extraction, source tagging, conflict detection, and credibility initialization, then generates corresponding semantic representations and writes them into relational tables, semantic indexes, and relational association indexes, respectively. Field standardization converts time, drug name, task name, unit, and status values ​​from different sources into a unified format; key entity extraction identifies labels for entities such as users, drugs, diseases, family members, caregivers, equipment, and locations; source tagging records the information source; conflict detection determines whether new and old fact values ​​under the same user and the same fact key contradict each other; and credibility initialization assigns initial credibility based on source weight, collection time, and manual confirmation. For example, "Take blood pressure medication at 8 PM tonight" and "Take Norvasc at 8 PM" can be standardized to the same medication time and corresponding drug entity label, and written under the same fact key for subsequent arbitration.

[0040] Furthermore, the long-term memory engine includes tables that hierarchically store factual memory, preference memory, intervention memory, and effect memory. The fields of the factual memory table include user identifier, fact key name, fact value (the fact value corresponding to the fact key name, such as new and old fact values), source, credibility (determined based on source weight, collection time, and manual confirmation), and update time. The fields of the preference memory table include preference type, preference style, applicable time range, and credibility. The fields of the intervention memory table include trigger type, action type, result label, and failure reason. The fields of the effect memory table include action type, scene label, time bin, acceptance rate, aversion score, decay factor, and recovery period.

[0041] The above preference types may include communication preferences, time preferences, reminder frequency preferences, intervention channel preferences, family participation preferences, etc.; preference styles may include direct reminder type, care and guidance type, brief confirmation type, explanation type, etc.; applicable time range refers to the time conditions under which the preference takes effect, such as evening, 30 minutes before taking medication, rehabilitation training period, or the most recent 30 days.

[0042] The trigger types mentioned above may include medication not confirmed, insufficient water intake, incomplete exercise, abnormal sleep, low mood, device alarm, etc.; action types may include voice reminders, text reminders, care inquiries, family notifications, caregiver review, etc.; result labels may include accepted, completed, delayed completion, rejected, no response, escalation, etc.; failure reasons may include user rejection, not heard, inappropriate timing, content mismatch, device malfunction, etc.

[0043] The "Action Type" mentioned above has the same meaning as the action type in the intervention memory table above, referring to the category of intervention actions actually executed or planned to be executed by the system, such as voice reminders, text reminders, care inquiries, family notifications, and caregiver verification. Acceptance rate can be calculated based on "Number of interventions accepted or completed by the user / Total number of interventions" in the statistics window; aversion score can be obtained by weighting information such as user rejection, negative verbal feedback, closing reminders, requests to stop reminders, and negative records from family members or caregivers, with a value that can be set to 0-1 or 0-100. Recovery period refers to the time and number of positive closed-loop feedback cycles required to gradually reduce the impact of a certain type of negative feedback or highly disruptive state, such as 7 days, 14 days, or recovery of recommendation weight after 3 consecutive positive feedback cycles.

[0044] In specific embodiments, the long-term memory engine can adopt different storage methods based on data types. For basic fields that are easy to retrieve and sort, such as memory item identifier, user identifier, event identifier, fact key name, fact value, source type, update time, credibility, risk level, scene tag, and status identifier, they are preferably stored in a relational database or SQLite table. For complex semantic content that needs to retain context, such as the user's original expression, event summary, emotion judgment criteria, reasons for intervention failure, family supplementary explanations, caregiver notes, and execution evidence package details, they are preferably attached and saved as JSON documents. For the relationships between memory items, such as the same user, the same medication, the same care task, similar events, intervention ticket references, execution evidence package write-back, family confirmation, caregiver review, and support, conflict, substitution, or version inheritance relationships, they can be further stored in a graph database or key-value index.

[0045] Long-term memories (historical memory items or historical memories) in the long-term memory engine can be stored using structured databases, key-value stores, graph databases, or combinations thereof, and event type indexes, scene tag indexes, and time tag indexes can be created for memory items.

[0046] The long-term memory engine also includes a hot memory layer, a steady-state memory layer, and an archived memory layer. The hot memory layer is used to store recently frequently called and currently effective memory items. The steady-state memory layer is used to store currently effective fact memories, preference memories, intervention memories, and effect memories. The archived memory layer is used to store low-frequency historical events, expired memories, and versioned execution evidence packages, and can perform segmented archiving according to time windows.

[0047] In a specific embodiment, the hot memory layer is used to store recently frequently invoked (invoking more than a preset number) and currently effective memory items, such as the user's frequently triggered medication times in the past week, the currently executing rehabilitation plan, and the most recently verified reminder preferences, which can preferably be stored in a key-value cache; the steady-state memory layer is used to store currently effective factual memories, preference memories, intervention memories, and effect memories, such as the user's long-term disease history, long-term dietary restrictions, stable preferences, commonly used intervention methods, and their statistical effects, which can preferably be stored in a relational database and a JSON document library; the archived memory layer is used to store low-frequency historical events, expired memories, and versioned execution evidence, such as low-frequency anomalies from several months ago, expired medication plans, falsified preference records, and execution evidence formed under old strategy versions, which can be archived in segments according to time windows. This system (system backend) can determine the migration of memory items between the hot memory layer, steady-state memory layer, and archived memory layer based on the call frequency, the most recent verification time, the credibility, and whether it is still in the current care stage. The system backend can also periodically perform memory compression and merging tasks, merge memory items with similar content, consistent sources and consistent closed-loop verification results, freeze or split memory items with long-term conflicts and incomplete verification, and demote or archive memory items that exceed the expiration threshold and have not been verified recently, in order to control the size of memory and maintain retrieval efficiency.

[0048] In some embodiments, a memory maintenance module is further included for maintaining the memory data (long-term memory) of the long-term memory engine, wherein the memory maintenance includes: Failure Scan: The system iterates through memory items in the steady-state memory layer that have exceeded their preset validity period and have not been recently verified in a closed loop, updating their status based on the review results. Specifically, the system reads the last update time, most recent verification time, and preset validity period of each memory item, and checks whether there is a confirmation record consistent with that memory item in the recently executed evidence package. If no confirmation record is found, a low-intrusion review can be initiated to the user, family member, or caregiver. If the review fails or the review result is inconsistent with the original memory, the credibility of the memory item is reduced to below the deweighting threshold and moved to the archived memory layer; if the review result is consistent, its most recent verification time is refreshed and it is retained in the steady-state memory layer.

[0049] Conflict Detection: This function identifies multiple contradictory memory entries from different sources under the same user and fact key, and determines their retention or verification status based on priority scores. Priority scores are determined by both source weight and the most recent verification time. Sources such as manual confirmation, on-site records by caregivers, or family confirmation have higher weights than unconfirmed single voice statements or system inferences. When source weights are equal or close, memory entries with more recent closed-loop verification times have higher priority scores. For example, if one memory states "medication taken at 8:00 PM" and was confirmed by a caregiver on the same day, while another memory states "medication taken at 9:00 PM" and originated from a system inference a week prior, the former can be retained as a high-priority memory entry, while the latter can be marked as pending verification.

[0050] Merge and Compress: Multiple memories whose summary vectors have a cosine similarity exceeding a preset merging threshold, share the same source, and have consistent closed-loop verification results are merged, retaining the latest timestamp and the highest credibility score. The highest credibility score refers to the highest overall credibility score among the memories to be merged, which can be determined by source weight, time validity, verification consistency, and conflict penalty. After merging, the system can also write the source, evidence citations, and historical version numbers of the merged memories into the merge record for subsequent auditing and backtracking.

[0051] Hot migration: Promote steady-state memory item copies that have been called more frequently than the hot threshold to the hot memory layer cache, remove memory items that have not been called for several consecutive cycles from the hot memory layer cache and retain only steady-state layer copies.

[0052] In this embodiment, closed-loop verification refers to the process by which the system verifies whether the intervention has been accepted, completed, and effective by using feedback evidence such as user feedback, task completion records, changes in equipment status, family member confirmation, and caregiver confirmation after the intervention action is executed.

[0053] The memory credibility arbitration module calculates a comprehensive credibility score for activated memory items and removes memory items that are below the threshold.

[0054] In this embodiment, the memory arbitration module is used to handle conflicts, invalidations, and changes in credibility in long-term memory. For example, when there are inconsistencies between user dictation information, family member supplementary information, and historical care records, the system does not directly overwrite any of the information. Instead, it arbitrates credibility based on source type, temporal proximity, historical verification results, and task risk level. Preferably, when a high-risk task involves conflicting memories, the system automatically switches from executive intervention to confirmatory intervention, first obtaining new confirmatory information before deciding whether to execute subsequent actions. Thus, long-term memory is no longer a static database but becomes a decision-making basis with credibility stratification and a dynamic effectiveness mechanism.

[0055] The memory arbitration module employs a combination of rule-based scoring and model-based scoring. First, basic weights are assigned to different sources based on their level; for example, user input, family confirmation, caregiver records, and device-perceived results each have different initial weights. Then, a weighted adjustment is made based on time validity (time proximity), recent closed-loop verification consistency, and conflict penalty items corresponding to the source, resulting in a comprehensive credibility score for the memory item. The expression for this score is as follows: ; in, , , , These are the weighted values ​​for source weight, time validity, verification consistency, and conflict penalty, respectively. These weights can be set according to the actual situation.

[0056] In a specific embodiment, the source weight can be set to 0.85 for user input, 0.90 for family confirmation, 0.80 for caregiver record, and 0.75 for device perception result; the time validity can be 0-1, with higher values ​​for those closer to the present; the verification consistency can be 0-1, with higher values ​​for consistent evidence from multiple sources; the conflict penalty can be 0-0.3, deducting the overall credibility when information from different sources conflicts.

[0057] When the overall credibility of the activated memory item is lower than the execution threshold and the risk level is a high-risk scenario and / or a conflict scenario, the action-type intervention to be executed will be automatically switched to a confirmation-type intervention, and a secondary confirmation request will be sent to the user, family member, or caregiver.

[0058] The intervention ticket generation and distribution module, based on the effect memory stored in the long-term memory engine, generates personalized intervention plans consisting of multiple intervention tickets for the active memory items that have not been removed, and selects the intervention ticket with the highest priority for distribution and execution. The priority can be determined by weighting risk level (a quantitative score of 0-100, used to represent the degree of risk that the current care task may be not completed in time; low risk, medium risk, medium-high risk, and high risk are the segment results of this risk level), historical acceptance rate, expected completion rate (which can be calculated based on the historical completion situation under the same user, similar event type, same time bucket, and same intervention method, or can be weighted and predicted by combining recent acceptance rate, task difficulty, current emotional label, and time urgency), aversion score, time decay score, credibility score, and manual review requirements (0 for no review, 0.5 for recommended manual spot checks, and 1 for mandatory manual confirmation; or a score of 0-100 to represent the review intensity, and this score is included in the intervention ticket priority calculation). For example, the task risk score for a medication reminder event is 85, the urgency score is 0.90, and the aversion score for direct reminders in the evening is 0.78. The historical acceptance rate of "show concern first, then remind and require confirmation" is 0.82, which is higher than the 0.55 of "immediate strong reminder". Therefore, the system prioritizes the former.

[0059] When the overall credibility of activated memory items is below the execution threshold, and high-risk or conflicting scenarios exist, the system can switch to a pending confirmation state. High-risk scenarios include missed medication, suspected falls, prolonged unresponsiveness, abnormal vital signs, going out at night, continuous negative emotions, or emergency calls for help; conflicting scenarios include contradictory memories under the same fact key, inconsistencies between user feedback and family confirmation, inconsistencies between device perception results and manual records, and conflicts between current events and historical preferences. When entering the above scenarios, the system will preferentially issue a confirmatory intervention first, or notify the family or caregiver to verify the information.

[0060] In this embodiment, the intervention ticket generation and distribution module generates personalized intervention plans based on structured care event descriptions and activated memory items. Preferably, the personalized intervention plan is not a single-sentence response, but an intervention ticket.

[0061] In some embodiments, the intervention ticket includes: a current event identifier (standard caregiver event object identifier), the key memory item invoked, the recommended intervention method, the expected execution timing, the expected feedback indicators, recurrence triggering conditions, the escalation threshold, and the review requirements. The invoked key memory item is preferably an active memory item that has not been removed after processing by the memory credibility arbitration module, and the corresponding memory item identifier, comprehensive credibility score, and fields involved in the decision-making can be recorded in the ticket. Through the intervention ticket, the system can explicitly record "why this intervention method was chosen, what result is desired, and under what conditions should it be escalated or reverted." Recommended intervention methods include voice reminders, text reminders, care inquiries, family notifications, and caregiver review; the expected execution timing can be determined based on the task deadline, user preference time, and time urgency; expected feedback indicators include acceptance, completion, completion delay, and negative reactions; recurrence triggering conditions include no response, incomplete, and recurrence of abnormality; escalation thresholds include risk score, number of no-response instances, and timeout length; and review requirements include no review required, review recommended, and review mandatory.

[0062] Furthermore, intervention tickets are stored using structured JSON objects or relational records, allowing the system to convert personalized suggestions from the natural language layer into executable, traceable, and rewritable intermediate control objects.

[0063] The intervention ticket generation and distribution module generates multiple candidate intervention tickets around the same standard caregiver event and compares them based on effect memory. This comparison is based not only on current user needs but also on historical acceptance rates, level of disruption, failure rates, and risk levels in similar scenarios. For example, for medication reminder events, it compares options such as "direct reminder," "care first, then reminder," "remind and request confirmation," and "delayed reminder after failure," prioritizing the intervention method with better historical results in the current time period and emotional state. Through this mechanism, the system truly embodies the technical characteristic of "long-term memory participating in personalized intervention."

[0064] Furthermore, the long-term memory engine updates the effect memory parameters of the activated memory items corresponding to the intervention ticket based on the feedback information continuously received by the closed-loop feedback module after the intervention ticket is executed. The effect memory parameters describe the dynamic effectiveness of a particular intervention method under specific user, scenario, and time buckets, including initial acceptance, recent acceptance, decay parameter, aversion threshold, recovery period, and scenario adaptation weight. Specifically, initial acceptance represents the probability of acceptance during the first or early execution; recent acceptance represents the actual acceptance within the most recent time window; decay parameter represents the rate at which historical effects decrease over time; aversion threshold represents the threshold at which negative feedback accumulates to the point where the intervention method needs to be suppressed; recovery period represents the observation time before the suppressed intervention method re-enters the candidate set; and scenario adaptation weight represents the applicability of the intervention method under different scenario labels.

[0065] The update process employs a combination of incremental statistics and time decay. The steps include: first, performing bucketed statistics on activated memory items by time period, scene label, and emotion label; then, exponentially smoothing the acceptance rate, completion rate, and aversion score of the activated memory items. For example, when the intervention is accepted, the corresponding accept_rate is increased; when the intervention triggers significant resistance or repeated prompting fails, the dislike_score is increased and the decay_factor is accelerated; when a certain type of intervention is not triggered for a period of time or its effect is verified to have recovered, its recommendable weight is gradually restored based on the recovery_cycle.

[0066] Through the above mechanism, the system can continuously maintain a long-term model of "the effectiveness of intervention changes over time" with low computational cost.

[0067] The closed-loop feedback module forms an execution evidence package from the continuously received feedback information after the intervention ticket is executed, and writes the execution evidence package back to the long-term memory engine and the policy version record table respectively.

[0068] In some embodiments, the evidence package includes the corresponding intervention ticket identifier, execution time, user immediate response, acceptance, completion, completion delay, need for further reminder, negative reaction, escalation of treatment, manual confirmation result, and final effect label; feedback information includes user verbal feedback, changes in environmental status, care task completion records, family confirmation, caregiver confirmation, and / or periodic review results.

[0069] The aforementioned escalation measures refer to upgrading the intervention level of current caregiving events or incomplete care tasks. For example, this could be escalating from a simple reminder to a repeated reminder, requiring confirmation, notifying family members, notifying caregivers, manual intervention, or emergency intervention. Manual confirmation results may include confirmation of completion, confirmation of incompleteness, confirmation of false alarm, confirmation of continued observation, or confirmation of escalation required. The final outcome label may include effective, partially effective, ineffective, causing negative reactions, or requiring escalation. The final outcome can be comprehensively confirmed by combining manual confirmation results with user feedback, equipment status, and task completion records, and is not limited to manual confirmation. Periodic review results may include whether care tasks are continuously completed, whether risks have been eliminated, whether user condition has improved, whether negative feedback has decreased, whether the intervention method needs adjustment, and whether further manual intervention is needed.

[0070] In some embodiments, the closed-loop feedback module collects feedback data through a local message queue, task callback interface, or HTTP / HTTPS return interface, and encapsulates it into a unified execution evidence package JSON. After the execution evidence package is generated, it is written to a subset of intervention memory and effect memory in long-term memory, and also to a policy version record table to support post-event auditing, failure tracking, and version rollback.

[0071] Furthermore, for high-risk scenarios, the closed-loop feedback module preferably incorporates a status grid escalation mechanism. A status grid includes at least the initial event status, pending confirmation status, confirmed but incomplete status, persistent abnormal status, escalated handling status, and archived status. The system transitions between these statuses based on feedback evidence, including user voice or text responses, click confirmations, task completion records, equipment status changes, caregiver verification, family confirmation, records of unresponsive actions within timeouts, and continuous abnormality detection results—information that demonstrates the intervention execution status and changes in risk.

[0072] When new closed-loop evidence meets the escalation criteria, the system raises the intervention level. Escalation criteria may include failure to complete the task beyond a preset deadline, lack of response to multiple reminders, reaching a high-risk level, escalating negative emotions, persistently abnormal device data, confirmation of risk through manual review, or repeated occurrences of the same event within a short period. The intervention level indicates the intensity of the intervention and the scope of participation, and may include low-intrusion prompts, repeated reminders, requiring user confirmation, notifying family members or caregivers for review, escalation to manual intervention or emergency intervention, etc. If the evidence shows the risk has been eliminated, the system promptly reverts to its previous state or terminates the escalation. In this way, closed-loop feedback not only records the results but also directly participates in subsequent status control.

[0073] The closed-loop feedback mechanism in this embodiment does not rely on chassis movement, but emphasizes that "the intervention results must be returned to the memory and strategy update link".

[0074] The self-evolutionary update module triggers event-level real-time updates and periodic-level batch updates based on the execution evidence package.

[0075] In this embodiment, the self-evolutionary update module adopts a dual-channel approach of "event-level real-time update + periodic-level batch update". The event-level real-time update is executed immediately after the end of a single intervention loop and is used to update the credibility of the memory items associated with the current event (standard caregiver event object), the intervention result label, and the effect parameters. The periodic-level batch update is triggered by a scheduled task or background service on a daily, weekly, or preset basis. It aggregates and analyzes the execution evidence package within a certain time window, generates a new set of strategy parameters and a version number, and writes them into the strategy configuration table or strategy configuration file for direct use by the subsequent intervention generation module.

[0076] In some embodiments, the self-evolutionary update module includes a micro-evolutionary layer and a macro-evolutionary layer. The micro-evolutionary layer is used to locally update the memory items and effect parameters involved in a single event after its conclusion. For example, if a user shows significant resistance to direct reminders in the evening but accepts caring reminders, the system immediately lowers the acceptability parameter of "direct reminders in the evening" and increases the weight of "caring reminders in the evening." If a factual memory is falsified in this closed loop, its credibility is lowered or it is moved to a pending verification state.

[0077] The macro-level self-evolution layer is used to optimize system-level parameters overall after multiple rounds of events or preset cycles. System-level parameters include intervention selection weights, risk thresholds for different scenarios, repeat reminder limits, escalation trigger conditions, memory expiration cycles, and effect decay model parameters. Intervention selection weights refer to the parameters assigned to different evaluation factors when generating candidate intervention tickets, such as historical acceptance weights, completion rate weights, risk urgency weights, aversion penalty weights, credibility weights, and manual review cost weights. The macro-level self-evolution layer is not limited to modifying prompts; instead, it continuously fine-tunes the above weights and related system-level parameters based on extensive closed-loop evidence, making the system more inclined to invoke trusted memories and select intervention methods with high success rates and low disruption levels under different users and scenarios.

[0078] The self-evolutionary update module also preferably incorporates a memory lifecycle management mechanism. The memory lifecycle includes six states: generation, activation, verification, demotion, freezing, and archiving. When a memory item remains unverified for an extended period, repeatedly conflicts with closed-loop feedback, or is no longer applicable at the current stage, the self-evolutionary update module automatically reduces its validity weight, marks it as pending verification, or archives it. For example, a memory might record a user's preference for "reminding medication at 8:00 PM," but multiple closed-loop feedbacks show the user refused at 8:00 PM and completed the reminder at 9:00 PM; or a memory might record a rehabilitation exercise suitable for the current stage, but the caregiver's review shows the exercise is no longer applicable due to changes in the recovery stage. These can all be considered conflicts between the memory and closed-loop feedback.

[0079] When the effectiveness of a certain type of intervention continues to deteriorate, the self-evolutionary update module automatically reduces its recommended scope or marks it as a high-intrusion method. The effectiveness of an intervention can be evaluated using indicators such as acceptance rate, task completion rate, completion delay, number of repeated reminders, negative feedback rate, aversion score, escalation trigger rate, manual review pass rate, relapse rate, and long-term user satisfaction. If a certain type of intervention shows a decrease in acceptance rate, an increase in negative feedback rate, or an abnormal increase in escalation trigger rate within a continuous statistical window, the system reduces its recommended scope or marks it as a high-intrusion method. Through the above mechanism, the system can avoid disordered expansion of long-term memory and accumulation of errors.

[0080] The effectiveness of a particular intervention program can be primarily assessed using acceptance rate, task completion rate, negative feedback rate, aversion score, and escalation trigger rate. For example, if the acceptance rate decreases, the negative feedback rate increases, or the escalation trigger rate increases within a consecutive statistical window, the intervention program is considered to have deteriorated in effectiveness.

[0081] In some embodiments, the system also sets up a policy version record table, which includes at least the policy version number, generation time window, applicable scenario, negative feedback rate, upgrade trigger rate, and rollback flag. After a periodic batch update generates a new policy version, the old policy is not immediately and permanently overwritten. Instead, the negative feedback rate and success rate corresponding to the new policy version are statistically analyzed in subsequent time windows. When the new policy version causes the negative feedback rate to rise continuously or the upgrade trigger rate to rise abnormally, the system automatically performs a version rollback or tightens the corresponding risk threshold. By introducing a policy version record table and a rollback control mechanism, self-evolutionary updates are no longer just abstract optimizations, but become an implementable technical process with version management and risk mitigation capabilities.

[0082] The specific working process of this system is as follows: If the system identifies a "medication reminder event," it first retrieves the medication name, reminder time, and past execution records from factual memory; from preference memory, it extracts the user's acceptance preference for reminder tone and confirmation methods; from effect memory, it extracts the user's historical acceptance of different reminder methods during that time period; and from intervention memory, it extracts the success rate and reasons for failure of similar reminders in the past. Based on the above information, the system generates multiple candidate intervention tickets and selects the optimal ticket for execution. After execution, it generates an execution evidence package based on whether the user responded, whether the medication was actually taken, whether a second reminder is needed, and whether resistance was triggered, and updates the parameters of the relevant memories and effect memories accordingly. In this way, the system completes a complete cycle of "memory retrieval—intervention execution—closed-loop feedback—self-evolutionary update."

[0083] In some embodiments, the intelligent agent for elderly care is built using a layered framework consisting of an "interaction access layer—cognitive decision-making layer—memory engine layer—execution verification layer—self-evolution layer." The interaction access layer receives voice, text, interface input, sensor events, and remote business messages; the cognitive decision-making layer performs event parsing, task identification, risk classification, and intervention planning; the memory engine layer performs long-term memory writing, retrieval, arbitration, merging, and archiving; the execution verification layer performs reminders, confirmations, reviews, and result collection; and the self-evolution layer performs event-level updates, periodic aggregation analysis, and strategy version updates. Each layer communicates its state through a shared task context object, which includes an event object identifier, an active memory set, a candidate intervention ticket set, the current execution state, an evidence package identifier, and a strategy version number. Through this layered framework, the intelligent agent no longer directly generates the final result as a single model, but rather forms a decomposable, verifiable, and rewritable closed-loop processing chain.

[0084] Furthermore, the cognitive decision-making layer can adopt a collaborative approach between a master control agent and internally specialized sub-agents. The system utilizes the master control agent to uniformly control task routing and context state, while event parsing sub-agents, memory orchestration sub-agents, intervention planning sub-agents, verification sub-agents, and self-evolving sub-agents each undertake different functions. They collaborate with fixed-field data objects through standardized task interfaces, thereby improving the processing stability and module verifiability in elderly care scenarios.

[0085] In a specific embodiment, the intelligent agent system can be built based on a Python language environment, preferably using LangChain as the intelligent agent orchestration framework and Flask as the service carrying framework. The driving model is preferably the Alibaba qwen3.6-plus large language model. LangChain is used to implement tool calls, task orchestration, context passing, and result writing between the master intelligent agent and each sub-intelligent agent; Flask is used to carry event access, status query, and result writing services via an HTTP interface; qwen3.6-plus is used for event understanding, memory retrieval decision-making, intervention ticket generation, and feedback evidence analysis; in scenarios requiring access to asynchronous device events, a message subscription mechanism can also be used to trigger external events.

[0086] It should be noted that indicators or terms not specifically described in this embodiment may be consistent with existing technologies in the field of elderly care and companionship, and will not be repeated here.

[0087] In summary, this system effectively reduces the heterogeneous coupling problem between different data sources by unifying voice content, family member recordings, nursing records, task logs, and exception messages into standard event inputs through a unified JSON event object, a multi-source interface access mechanism, and a structured event parsing method. Compared to solutions that rely solely on a single dialogue input or distributed log recording, this invention improves the elderly care system's ability to access multi-source information and the consistency of event processing, providing a unified data foundation for subsequent memory retrieval and intervention decisions.

[0088] By constructing a hierarchical long-term memory system comprising factual memory, preference memory, intervention memory, and effect memory, and combining it with relational storage, JSON document storage, and associated indexing mechanisms, the system can simultaneously achieve precise retrieval of memory items, historical tracing, and dynamic updates. Compared to solutions that only store single-round contexts or static user profiles, this invention significantly improves information retention capabilities, scene retrieval efficiency, and individualized matching accuracy during long-term care.

[0089] By introducing a memory arbitration mechanism comprised of source weighting, time validity, verification consistency, and conflict penalties, long-term memory is transformed from a static, accumulated information repository into a dynamic decision-making basis with credibility assessment capabilities. Especially in medication reminders, anomaly alerts, and other high-risk care scenarios, the system can determine the reliability of relevant memories before execution and automatically switch to confirmatory intervention if credibility is insufficient. This improves intervention safety, reduces the probability of false triggers, and mitigates the risks associated with the long-term propagation of erroneous memories.

[0090] By setting up intervention tickets, effect memory, and execution evidence packages, each personalized intervention has a complete chain of "traceable generation basis, recordable execution process, and rewritable execution results." Specifically, intervention tickets transform natural language suggestions into structured intermediate objects, and execution evidence packages uniformly feed user responses, completion status, and manual confirmation results back to the memory and strategy module, thereby significantly enhancing the verifiability, auditability, and targeted optimization of elderly care tasks.

[0091] By continuously updating acceptance rate, success rate, aversion score, time decay parameters, and recovery cycle, the system can gradually learn "which intervention method is more effective under what time, scenario, and emotional conditions." Therefore, the system can proactively reduce intervention methods with low acceptance, high intrusiveness, or high historical failure rates, thereby improving reminder success rate, reducing user resistance, and enhancing the long-term care experience.

[0092] Through a dual-channel self-evolution mechanism of "event-level real-time updates + periodic-level batch updates," the system can quickly correct local memory and local effect parameters after a single event loop, and continuously adjust intervention selection weights, risk thresholds, repeated reminder limits, and memory expiration rules over a longer time scale. Therefore, this invention can significantly reduce manual maintenance costs and improve the system's stability, continuous adaptability, and strategy evolution capabilities during long-term operation.

[0093] The above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, characterized in that, include: The multi-source event access module encapsulates the received multi-source care data into standard care event objects; The structured event parsing module performs event classification, risk grading, emotion discrimination, time binning, and scene labeling on the standard caregiver event objects to obtain the parsing results; The long-term memory retrieval module performs structured index retrieval, semantic representation retrieval, and relation expansion retrieval in the long-term memory engine based on the structured fields corresponding to the parsing results, the summary of the standard care event object, historical intervention tickets, key entity tags extracted from the standard care event object, and known event relationships maintained by the long-term memory engine. The retrieval results are then fused and rearranged to obtain the activated memory item of the standard care event object. The memory credibility arbitration module calculates the comprehensive credibility score of the activated memory item based on the source weight, time validity, verification consistency, and conflict penalty item, and removes memory items that are below the threshold. The intervention ticket generation and distribution module generates a personalized intervention plan consisting of multiple intervention tickets based on the effect memory stored in the long-term memory engine for the active memory items that have not been removed, and selects the intervention ticket with the highest priority to distribute and execute. The closed-loop feedback module forms an execution evidence package from the continuously received feedback information after the intervention ticket is executed, and writes the execution evidence package back to the long-term memory engine and the strategy version record table respectively. The self-evolutionary update module triggers event-level real-time updates and periodic-level batch updates based on the execution evidence package. The intervention ticket includes a standard caregiver event object identifier, the key memory item invoked, the recommended intervention method, the expected execution timing, the expected feedback indicators, the recurrence triggering conditions, the escalation threshold, and the review requirements; The execution evidence package includes the corresponding intervention ticket identifier, execution time, user's immediate response, whether it was accepted, whether it was completed, completion delay, whether a reminder is needed, whether a negative reaction occurred, whether it entered escalation processing, manual confirmation result, and final effect label; The feedback information includes user language feedback, changes in environmental status, records of care task completion, family member confirmation, caregiver confirmation, and / or periodic review results; The long-term memory engine includes tables that hierarchically store fact memory, preference memory, intervention memory, and effect memory. The fields of the fact memory table include user identifier, fact key name, fact value, source, credibility, and update time. The fields of the preference memory table include preference type, preference style, applicable time range, and credibility. The fields of the intervention memory table include trigger type, action type, result label, and failure reason. The fields of the effect memory table include action type, scene label, time bucket, acceptance rate, aversion score, decay factor, and recovery period. The long-term memory engine also includes a hot memory layer, a steady-state memory layer, and an archive memory layer; wherein, the hot memory layer is used to store recently frequently called and currently effective memory items, the steady-state memory layer is used to store currently effective fact memory, preference memory, intervention memory, and effect memory, and the archive memory layer is used to store low-frequency historical events, expired memories, and versioned execution evidence, and performs segmented archiving according to time windows.

2. The personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution as described in claim 1, characterized in that, For the newly added care information corresponding to the standard care event object, the long-term memory engine performs field standardization, key entity extraction, source marking, conflict detection and credibility initialization, and then generates the corresponding semantic representation and writes it into the relational table, semantic index and relation association index respectively.

3. The personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution as described in claim 1, characterized in that, The semantic representation retrieval includes: The summary text of each memory item is encoded as a dense vector representation of a fixed dimension, and the vector is written into a vector index library; the summary of the standard care event object is encoded as a dense vector, and the most semantically relevant memory item is recalled from the vector index library as the retrieval result corresponding to the semantic representation retrieval.

4. The personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution as described in claim 1, characterized in that, The search results are merged and rearranged, including: The comprehensive score of each historical memory item in the search results is calculated based on the structural matching score, semantic similarity score, relational association score, credibility score, and time decay score corresponding to the historical memory item; based on the calculated comprehensive score of the historical memory item, historical memory items with a comprehensive score higher than a preset threshold are selected to constitute the activated memory items of the standard care event object; The structure matching score is determined based on the matching degree of event type and scene tag of each memory item in the search results. The semantic similarity score is determined based on the vector cosine similarity between the summary of the standard caregiver event object and the summary of each memory item in the search results. The relational association score is determined based on the overlap between the key entity tags involved in the standard caregiver event object and the key entity tags associated with each memory item in the search results. The credibility score is taken from the comprehensive credibility score of each memory item in the search results obtained through the memory credibility arbitration module. The timeliness decay score is calculated based on the interval between the last update time and the current time of each memory item in the search results using an exponential decay function.

5. A personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, as described in claim 1, is characterized in that... It also includes a memory maintenance module for maintaining the memory data of the long-term memory engine, the memory maintenance of which includes: Traverse the memory items in the steady-state memory layer that have exceeded the preset validity period and have not been recently verified in closed-loop, and update their status according to the verification results; Identify multiple memory items with different sources and contradictory content under the same user and the same fact key, and determine the retention or verification status based on priority scoring; For multiple memory items whose cosine similarity between summary vectors exceeds a preset merging threshold, have the same source, and have consistent closed-loop verification results, merge them while retaining the latest timestamp and the highest credibility. The steady-state memory item copy whose recent call frequency exceeds the hotness threshold is promoted to the hot memory layer cache. The memory items in the hot memory layer that have not been called for multiple consecutive cycles are removed from the cache and only the steady-state layer copy is retained.

6. The personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution as described in claim 1, characterized in that, The long-term memory engine updates the parameters of the effect memory of the activated memory item corresponding to the intervention ticket based on the feedback information continuously received by the closed-loop feedback module after the intervention ticket is executed; the parameters of the effect memory include initial acceptance, recent acceptance, decay parameter, aversion threshold, recovery period and scene adaptation weight. The update uses a combination of incremental statistics and time decay, including: First, the activated memory items are statistically analyzed by time period, scene tag, and emotion tag. Then, the acceptance rate, completion rate, and aversion score of the activated memory items are updated exponentially.

7. The personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution as described in claim 1, characterized in that, When the overall credibility of the activated memory item is lower than the execution threshold, and its risk level belongs to a high-risk scenario or a conflict scenario, the action-type intervention to be executed is automatically switched to a confirmation-type intervention, and a secondary confirmation request is sent to the user, family member, or caregiver.

8. A personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, as described in claim 1, is characterized in that, The self-evolutionary update module includes a micro-evolutionary layer and a macro-evolutionary layer. The micro-evolutionary layer is used to locally update the memory items involved and their effect memory parameters after a single intervention event ends. The macro-evolutionary layer is used to optimize the system-level parameters as a whole after multiple rounds of intervention events or a preset period. The system-level parameters include intervention selection weights, risk thresholds under different scenarios, repeated reminder limits, upgrade trigger conditions, memory expiration period, and effect decay model parameters.

9. A personalized intervention system for elderly care companionship based on long-term memory and closed-loop feedback self-evolution, as described in claim 1, is characterized in that, The self-evolutionary update module adopts a memory lifecycle management mechanism. When a memory item has not been verified for a long time, has repeatedly conflicted with closed-loop feedback, or is no longer applicable in the current stage, the self-evolutionary update module automatically reduces its effective weight or archives it. When the effect of a certain intervention scheme continues to deteriorate, the self-evolutionary update module automatically reduces its recommended scope or marks it as a high-disruption method. The memory lifecycle includes generation, activation, verification, demotion, freezing, and archiving.

Images