A Context Processing Method and Related Equipment for Home Smart Central Screen Devices Based on Member Awareness

By employing a context processing method for smart home screen devices with member awareness, the problems of memory confusion and context overflow during multi-member use are solved. This enables cross-device memory sharing and task routing, improving the memory recall rate and task processing efficiency of smart home screen devices, and meeting the requirements for privacy protection and data persistence.

CN123115336APending Publication Date: 2026-11-17CHINA TELECOM ARTIFICIAL INTELLIGENCE TECHNOLOGY (BEIJING) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611348291.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In multi-member scenarios, smart home screen devices suffer from issues such as memory confusion, context overflow, insufficient semantic awareness, isolated memory between the device and the cloud, and lack of task scheduling, resulting in low efficiency in memory sharing and task processing.

Method used

A context processing method for smart home screen devices based on member awareness is adopted. By identifying the current interactive member, an initial context vector is constructed. Combined with collaborative memory management on the device side and the cloud, context tokens are dynamically allocated, and differentiated compression and privacy processing are performed to achieve cross-device memory sharing and task routing.

Benefits of technology

It improves the memory recall rate of multiple family members, enhances the memory sharing capability across devices, optimizes the retention rate of context information, reduces the latency of task response on the device side, ensures privacy protection and data persistence, and solves the problem of memory data loss.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115336A_ABST
    Figure CN123115336A_ABST
Patent Text Reader

Abstract

This invention discloses a context processing method and related equipment for a home smart central screen device based on member awareness. The method includes: assessing the complexity and privacy classification of the current task of the home smart central screen device; calculating the edge-cloud execution score; determining whether the inference should be performed on the edge or in the cloud; allocating a context token budget to the executor; calculating the ISM importance score; performing differentiated compression on the context of the current task; combining the compressed context into structured prompts; performing inference at the executor's location; retrieving the memory of the interactive content in the inference result; and simultaneously writing key information to the edge and cloud after anonymization. This member-aware context processing method for home smart central screen devices significantly improves the memory recall rate of multiple family members, greatly enhances cross-device memory sharing capabilities, and significantly increases the effective information retention rate of semantically aware compression. This invention is applied in the field of text processing technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of text processing technology, and in particular to a context processing method and related equipment for a smart home screen device based on member perception. Background Technology

[0002] Currently, smart home screen devices can recognize and process text directly input by users or converted from voice, enabling intelligent response and control of user commands. Furthermore, a single smart home screen device can serve multiple users (e.g., multiple family members such as father, mother, grandfather, and grandmother), thus offering good economic benefits. However, current smart home screen devices suffer from the following technical problems: Problem 1: Confusion of long memories among multiple members: The lack of a memory isolation and sharing mechanism based on family members leads to mutual contamination of historical memories when multiple members share devices.

[0003] Question 2: Severely limited context window at the edge: The computing power of the chip in the home screen is limited (typical context window is 2K~4K tokens). Long conversations among multiple members cause the context to overflow rapidly, and key historical information is forcibly discarded.

[0004] Question 3: Lack of semantically aware context compression mechanism: Existing solutions either truncate after the fact or use a uniform summary rate, which cannot distinguish between key information (such as "the child is allergic to peanuts") and redundant information (such as "the weather is nice today").

[0005] Question 4: End-to-end memory silos: Pure end-side solutions lack persistent memory data and cross-device sharing capabilities; pure cloud solutions face privacy leakage risks and are insufficient in responding to real-time low-latency requirements.

[0006] Question 5: Lack of task scheduling: When faced with family interaction tasks of varying complexity (simple control commands vs. family decision-making discussions), there is a lack of intelligent edge-cloud task routing mechanisms, which means that even lightweight tasks need to go to the cloud, while heavyweight tasks are limited by edge capabilities. Summary of the Invention

[0007] To address at least one of the aforementioned technical problems, the present invention aims to provide a context processing method and related equipment for a home smart central screen device based on member perception.

[0008] On one hand, embodiments of the present invention include a context processing method for a home smart central screen device based on member awareness, the method comprising: Identify the currently interacting members of the smart home central display device; Collect the current scene context of the smart home central screen device, and construct an initial context vector based on the current scene context; Based on the initial context vector, a search is performed on the edge of the smart home screen device. When the hit rate of the edge memory determined by the search is lower than the hit rate threshold, a supplementary search is asynchronously initiated to the cloud. The search results of the edge and / or the cloud are comprehensively scored, and the Top-K memory is recalled and injected into the context based on the scoring results. The complexity and privacy classification of the current task of the smart home screen device are evaluated, the edge-cloud execution score is calculated, and it is determined whether the inference is performed on the edge or the cloud. A context token budget is allocated to the executor, where the executor is the edge or the cloud. When the cumulative number of tokens for the current task of the smart home screen device exceeds the dynamic compression threshold, the ISM importance score is calculated, and differential compression is performed on the context of the current task to obtain the compressed context. The compressed context, recalled memory, and current query are combined into structured prompts, and inference is performed at the executor. Obtain the reasoning result, perform memory extraction on the interactive content in the reasoning result, obtain key information that meets the writing conditions, perform desensitization processing on the key information, and synchronously write the key information to the terminal and the cloud.

[0009] Furthermore, the process of obtaining the lyrics text to be translated also includes: The key information is output in a multimodal linkage manner to update the context memory summary on the device side; and the incremental synchronization task in the cloud is triggered as needed.

[0010] Furthermore, the identification of the current interactive members of the smart home central screen device includes: The currently interacting member is identified through voiceprint, person, or device binding; Initiate the memory space for the currently interacting member and determine whether the currently interacting member is being used for the first time. If so, synchronize the historical memory of the currently interacting member.

[0011] Furthermore, the comprehensive scoring of the search results from the endpoint and / or the cloud includes: According to the formula Score(m,q,u) =α·Sim(Em, Eq)+β·Rel(m,u)·e^(-λΔt)+γ·Freq(m)+δ·Novelty(m,q) The comprehensive score Score(m,q,u) is calculated; where m is memory, q is query, u is user, Em is memory semantics, Eq is query semantics, Sim is cosine similarity; Rel(m,u) is member relevance weight; e^(-λΔt) is time decay; Freq(m) is access frequency; Novelty(m,q) is information novelty; α, β, γ, δ are weight coefficients, satisfying α+β+γ+δ=1.

[0012] Furthermore, the process of assessing the complexity and privacy of the current task of the smart home screen device, calculating the edge-cloud execution score, and determining whether the inference is performed on the edge or in the cloud includes: According to the formula RouteScore_cloud(q)=w_c·C(q)-w_p·Pri(q)+w_n·Net_score The decision score RouteScore_cloud(q) is calculated, where C(q) is the task complexity score, Pri(q) is the privacy level, Net_score is the overall score of the current network quality, and w_c, w_p, and w_n are weight coefficients. Set the routing threshold τ_route; If the decision score RouteScore_cloud(q) is greater than the routing threshold τ_route, the inference will be executed in the cloud; otherwise, the inference will be executed on the device.

[0013] Furthermore, the allocation of the context token budget to the executor includes: According to the formula B_alloc(u, θ) = B_total × [r(u) × g(θ)] / Σi[r(ui) × g(θi)] The context token budget B_alloc(u, θ) is calculated; where B_total is the total available token budget, u is the requesting member, θ is the current task type, r(u) is the role weight of member u, g(θ) is the complexity multiplier of task θ, r(ui) is the role weight of the i-th member u, and g(θi) is the complexity multiplier of the i-th task θ.

[0014] Furthermore, the calculation of the ISM importance score includes: According to the formula I(sk)=w1·E(sk)+w2·A(sk)+w3·N(sk)+w4·P(sk, u) The ISM importance score I(sk) is calculated, where sk is the k-th dialogue segment, E represents emotional intensity, A represents actionability, N represents entity novelty, P represents personalization gain, and w1, w2, w3, and w4 are weights.

[0015] Furthermore, the differential compression of the current task's context includes: According to the formula C_ratio(θ) = [B_avail B_mem B_cur] / B_hist · f(θ_task) The task type-aware adaptive compression ratio C_ratio(θ) is calculated; where B_avai is the total number of available context window tokens, B_mem is the number of tokens occupied by memory injection, B_cur is the real-time input of the current dialogue, B_hist is the total number of original tokens in the history dialogue, and f(θ_task) is the task-aware factor. Based on the task type-aware adaptive compression ratio C_ratio(θ), compression is performed on the context of the current task.

[0016] On the other hand, embodiments of the present invention also include a computer device, including a memory and a processor, the memory for storing at least one program, and the processor for loading at least one program to execute the member-aware home smart central screen device context processing method in the embodiments.

[0017] On the other hand, embodiments of the present invention also include a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the member-aware home smart central screen device context processing method in the embodiments.

[0018] The beneficial effects of this invention are as follows: The context processing method for home smart screen devices based on member perception in the embodiments significantly improves the memory recall rate of multiple family members, greatly enhances the cross-device memory sharing capability, significantly improves the effective information retention rate of semantic perception compression, and significantly reduces the occupancy of context tokens; the response latency of simple tasks on the device side is maintained at the sub-second level, and the inference quality of complex tasks in the cloud is improved; the PIE privacy desensitization mechanism ensures compliance with the Personal Information Protection Law, and the memory data is persisted through the cloud CCM, ensuring lossless memory migration after device replacement / fault recovery and solving the problem of data loss. Attached Figure Description

[0019] Figure 1This is a schematic diagram of an edge-cloud converged system architecture in which a context processing method based on member awareness for smart home screen devices can be applied in the embodiment. Figure 2 This is a schematic diagram of the differentiated lifecycle strategy in the embodiment; Figure 3 This is a schematic diagram illustrating the steps of the context processing method for a smart home central screen device based on member awareness in this embodiment. Figure 4 This is a flowchart illustrating the context processing method for a home smart central screen device based on member awareness, as shown in the embodiment. Figure 5 This is a schematic diagram of the three-layer family memory architecture in the embodiment; Figure 6 This is a flowchart of the CETR edge-cloud task routing decision-making process in the embodiment; Figure 7 This is a flowchart of the SACC semantic awareness context compression module in the embodiment; Figure 8 This is a timing diagram of a multi-family cloud collaborative interaction implementation example. Detailed Implementation

[0020] Terminology Explanation: 1. Home Memory-Conscious Agent (HMCA): A cloud-edge integrated multi-member memory management and context compression interaction system for smart home screens; 2. Personalized Layered Memory (PLM): This architecture divides family interactive memory into three layers: personal memory layer, family shared memory layer, and contextual memory layer. 3. Semantic-Aware Context Compression (SACC): An adaptive context compression mechanism based on dialogue segment importance scoring and task type awareness; 4. Dynamic Context Allocation Scheduler (DCAS): A scheduling algorithm that dynamically allocates context window token budgets based on task type and member roles; 5. Importance Scoring Module (ISM): A dialogue segment importance assessment model that integrates emotional intensity, actionability, and individual gains; 6. Family Home Screen (FHS): The core deployment terminal of this system, an intelligent device that integrates display, voice interaction and IoT control functions, and its product name is generally "Smart Screen" etc. 7. Episodic Memory (EPM): Short-cycle dialogue summaries organized according to family activity scenarios (dining / entertainment / learning / health); 8. Cloud Context Mirror (CCM): A cloud-maintained backup and synchronization service for long-term family memories, supporting cross-device memory sharing and persistence; 9. Cloud-Edge Task Router (CETR): A scheduling module that dynamically decides whether a task should be executed on the edge or in the cloud based on task complexity, privacy level, and network status. 10. Privacy Information Extractor (PIE): A security component that de-identifies / anonymizes privacy-sensitive fields before storing data in the cloud; 11. Member Context Router (MCR): The core routing module that identifies currently interacting members and switches the corresponding memory space.

[0021] Current smart home screen devices can adopt the following four technical approaches, and the principles and limitations of these technical approaches are as follows: Option A: Single dialogue model + session-level context The central screen in the home integrates a conversational AI engine, responding directly after receiving voice / touch commands. Drawbacks: 1) No cross-session memory: Each conversation is independent, and historical interactions cannot be automatically recalled in the next conversation; 2) No member distinction: The whole family shares the same context pool, and members' preferences contaminate each other; 3) Wasted context window: After multiple rounds of dialogue, redundant content accumulates and occupies a large amount of effective information space.

[0022] Option B: Static preference storage based on user profiles Basic preferences for each member are stored in a pre-configured manner. Drawback: 1) Static preferences cannot track dynamic evolution (such as changes in children's dietary preferences). 2) Lack of memory associations across activity scenarios and family-shared memory hierarchy.

[0023] Option C: Large-scale cloud model + lightweight home terminal (pure cloud solution) The terminal only handles wake-up and data acquisition; inference occurs entirely in the cloud. Drawbacks: 1) Privacy risks: All family conversations must be uploaded to the cloud, including highly sensitive health, parenting, and family financial information; 2) Latency issues: Cloud-based inference round-trip latency is typically 2-5 seconds, disrupting the smoothness of the conversation.

[0024] Option D: Limitations of the pure end-side large model scheme In recent years, solutions have emerged that allow for the complete execution of language models with a small number of parameters on the device side. Drawbacks: 1) The number of parameters in edge-side LLM is limited (usually ≤7B), resulting in insufficient complex reasoning ability and a significant decrease in response quality in high-demand scenarios such as family decision-making and health consultation; 2) The memory is stored only on the local machine. When the device is replaced or the device fails, all the stored data is lost, and there is no guarantee of persistence. 3) Cannot support cross-device memory sharing (the memory of the central screen in the home is completely separated from the mobile assistant and tablet). 4) Model updates require OTA updates for the entire device. Frequent updates affect user experience and consume a lot of data.

[0025] Based on the above principles, this embodiment provides a context processing method for home smart central screen devices based on member awareness. In this embodiment, the context processing method for home smart central screen devices based on member awareness can be applied to… Figure 1 The diagram shows an edge-cloud converged system architecture. (Refer to...) Figure 1 The edge-cloud converged system consists of two main parts: edge deployment and cloud services. The edge deployment has four layers, while the cloud service is an independent service layer. The specific structure is as follows: Cloud Service Layer (CL) Deployed in the public cloud, it connects to the edge FHS via a secure encrypted channel and includes the following sub-services: 1) Cloud-LLM service: Deploy large language models with ≥35B parameters to handle highly complex tasks (family decision-making, health consultation, complex workflow planning, etc.) routed from CETR. 2) Cloud Context Mirroring Service (CCM): Stores complete historical memories of all family members (after anonymization), provides cross-device memory retrieval and incremental synchronization services, and uses vector databases (such as Pinecone or self-built Faiss clusters) to support large-scale semantic retrieval; 3) Privacy Information Extraction Service (PIE): Performs NER named entity recognition on the memory writing CCM front end, and performs desensitization / replacement / encryption processing on high privacy fields such as name, ID number, mobile phone number, home address, and medical records, and outputs anonymized memory entries; 4) Family Memory Collaboration Service: Manages memory consistency collaboration among multiple FHS devices (living room central screen, bedroom central screen, kitchen small screen, etc.) under the same family account, using an eventual consistency protocol.

[0026] L4—Interactive Presentation Layer (Client-Side) Multimodal input acquisition (voice / touch / gesture), mid-screen card-style result display, IoT device control command issuance, and edge-cloud synchronized status indication.

[0027] L3—Core Processing Layer It contains the following five sub-modules, which are connected in series to form a pipeline of "perception → routing → allocation → compression → inference": Member Context Router (MCR): Identifies currently interacting members and remembers hot cache switching on the management side; Edge-Cloud Task Router (CETR): Assess task complexity and privacy level, and dynamically decide on the edge / cloud execution path (new module). Dynamic Context Allocation Scheduler (DCAS): Allocates token budgets based on member roles and task types; Semantic Context Compressor (SACC): Performs ISM scoring and adaptive compression; Lightweight LLM inference engine on the edge: It undertakes low-complexity tasks (daily control, simple queries, entertainment on demand, etc.) routed from CETR.

[0028] L2—Memory Management Layer The management side employs a three-layer hot memory cache (frequently accessed recent memory) and is responsible for incremental synchronization with the cloud-based CCM. Personal Hot Memory Cache (PLM_P_hot): A memory of frequently accessed personal preferences and events in the past 30 days, with a maximum capacity of 500 entries; Family shared hot memory cache (PLM_S_hot): Family shared memory for the past 30 days, maximum capacity 1000 entries; Contextual Memory Layer (EPM): A 24-hour rapidly decaying summary of the context, TTL=24h; Cloud mirror synchronization interface: Through the asynchronous incremental synchronization protocol, newly added / modified memory entries are periodically (triggered after each interaction) uploaded to the cloud CCM after being anonymized by PIE.

[0029] In this embodiment, refer to Figure 2 The memory management layer and cloud-based CCM can implement the following differentiated lifecycle strategies: Personal Hot Memory (PLM_P_hot): High-frequency data is retained on the device for 30 days; if there is no access for more than 30 days or the device capacity alarm is triggered, it will be automatically de-identified via PIE and archived to the cloud CCM (permanently stored, with retrieval weight decayed according to access frequency); the maximum device capacity is 500 records / person. Family Shared Hot Memory (PLM_S_hot): Retained on the device for 30 days; archived to the cloud CCM after the expiration period and retained for 365 days; automatically downgraded to "Family Log" after 365 days of inaccessibility (can be queried but not included in real-time retrieval); Episodic Memory (EPM): TTL=24h, after which it is automatically summarized as a 50-token keyword index and retained on the device for 30 days; it is uploaded to the cloud on demand (privacy-sensitive short-term activity records). Cloud-based CCM: Retained according to user privacy settings (up to 5 years), supports full deletion at any time; uses vector index + keyword dual index to accelerate retrieval.

[0030] L1—Sensing and Acquisition Layer (End Side) Microphone array, camera, environmental sensors (temperature, humidity / light), WiFi / BLE device detection, IoT device status synchronization interface, to collect multi-dimensional contextual data of the home environment in real time.

[0031] In this embodiment, refer to Figure 3 The context processing method for smart home screen devices based on member awareness includes the following steps: S1. Identify the currently interacting members of the smart home central screen device; S2. Collect the current scene context of the smart home screen device, and construct an initial context vector based on the current scene context; S3. Based on the initial context vector, perform a search on the edge of the smart home screen device. When the hit rate of the edge memory determined by the search is lower than the hit rate threshold, initiate a supplementary search to the cloud asynchronously, comprehensively score the search results on the edge and / or the cloud, and recall Top-K memories to inject context based on the score results. S4. Assess the complexity and privacy of the current task of the smart home screen device, calculate the edge-cloud execution score, determine whether the inference is performed on the edge or in the cloud, and allocate a context token budget to the executor. S5. When the cumulative number of tokens for the current task of the smart home central screen device exceeds the dynamic compression threshold, calculate the ISM importance score, perform differential compression on the context of the current task, and obtain the compressed context; S6. Combine the compressed context, recalled memory, and current query into structured prompts and perform inference at the executor; S7. Obtain the reasoning result, perform memory extraction on the interactive content in the reasoning result, obtain key information that meets the writing conditions, perform desensitization processing on the key information, and synchronously write the key information to the terminal and the cloud.

[0032] In this embodiment, the process of steps S1-S7 is as follows: Figure 4 As shown.

[0033] Reference Figure 4 When performing step S1, member awareness and context routing are performed: the current interactive member is identified through voiceprint / face / device binding, the MCR starts the memory space of the corresponding member (end-side hot memory cache + cloud CCM cold memory retrieval), and determines whether the current user is using it for the first time. If so, the member's historical memory is synchronized from the CCM.

[0034] Reference Figure 4 During step S2, context initialization is performed: collect the current scene context (time / space / family member status / IoT device data), construct the initial context vector C0; determine network availability, and initialize the end-to-cloud communication channel.

[0035] Reference Figure 4 When performing step S3, a cloud-edge collaborative hierarchical memory retrieval is performed: based on the initial context vector C0, a fast retrieval is first performed in the hot memory cache on the edge side (latency requirement ≤100ms); if the memory hit rate on the edge side is lower than the threshold θ_hit, a supplementary retrieval is initiated asynchronously to the cloud CCM; finally, the Top-K memory is recalled and injected into the context according to the comprehensive score of formula (1).

[0036] Specifically, in step S3, such as Figure 5 As shown, family interactive memory is divided into three layers (end-side hot cache) based on ownership scope and timeliness: 1) Personal Memory Layer PLM_P: Stored independently by family member ID, containing private information such as personal preferences, habits, and allergies, visible only to the individual. Hot data is cached on the device for 30 days, while earlier cold data is archived to the cloud CCM. 2) Family Shared Memory Layer PLM_S: Family-level shared information (schedule / rules / IoT scenario templates), with 30 days of hot data cached on the device side; 3) Contextual Memory Layer (EPM): Recent summaries organized according to activity scenarios, TTL=24h, stored on the device side, and transmitted to the cloud on demand.

[0037] End-to-cloud collaborative retrieval strategy: Prioritize fast vector retrieval performed in the hot cache on the end side; if the hit rate on the end side is lower than the threshold θ_hit (default 0.7), then asynchronously initiate supplementary retrieval to the cloud CCM, and take the Top-K after merging and sorting.

[0038] In step S3, the scoring formula used to calculate the overall score is as follows: Score(m, q, u) = α·Sim(Em, Eq) + β·Rel(m,u)·e^(-λΔt) + γ·Freq(m) + δ·Novelty(m,q) (1) Where m is memory, q is query, u is user, Sim is cosine similarity; Rel(m,u) is member relevance weight (Rel(m,u) is determined by the following rules: individual = 1.0, shared = 0.7, others = 0); e^(-λΔt) is time decay; Freq(m) is access frequency; Novelty(m,q) is information novelty; α+β+γ+δ=1.

[0039] In step S4, if it is decided that the inference will be executed on the client side, the executor is the client side; if it is decided that the inference will be executed in the cloud, the executor is the cloud. When executing step S4, refer to... Figure 6 The task routing and token budget allocation are performed on the edge and cloud: the CETR module evaluates the complexity and privacy of the current task, calculates the edge and cloud execution score according to formula (2), and decides whether the inference is executed on the edge lightweight LLM or the cloud large model; the context token budget is allocated simultaneously according to DCAS formula (3).

[0040] In step S4, the Edge-Cloud Task Routing Decision (CETR) algorithm is used to dynamically decide whether to execute the current task on the edge LLM or the cloud-based large model based on the task complexity C(q), privacy level Pri(q), and network quality Net_score. The edge-cloud execution score formula is as follows: RouteScore_cloud(q) = w_c·C(q) w_p·Pri(q) + w_n·Net_score (2) Higher scores lead to cloud-based routing, while lower / negative scores result in edge-based decision-making. Specifically, the following decision rule can be set: if RouteScore_cloud(q) > τ_route, route to the large cloud model; otherwise, execute a lightweight LLM on the edge. The meanings of each symbol are as follows: C(q): Task complexity score – estimates the reasoning depth of the task based on intent recognition and the number of entities, with values ​​[0,1]; simple instructions (turn off lights / play music / check the weather) C≈0.1~0.2; complex family decisions (travel planning / school district analysis / family health assessment) C≈0.7~1.0; Pri(q): Privacy level – weight of privacy fields involved in the task, with values ​​[0,1]; Pri≈0.7~1.0 for tasks containing family member names + precise addresses + medical records (biased towards client-side processing); Pri≈0.0~0.2 for general knowledge queries; Net_score: Current overall network quality score (bandwidth × availability × latency reciprocal normalization), with a value of [0,1]; when the network is disconnected, Net_score=0, forcing execution on the end side; w_c, w_p, w_n: weighting coefficients (e.g., w_c=0.5, w_p=0.4, w_n=0.1). These prioritize privacy protection to ensure that high-privacy tasks are not easily routed to the cloud simply because of good network quality. τ_route: Routing threshold, default 0.3, which can be adjusted by the user in the privacy settings (prefer privacy protection → increase τ; prefer reasoning ability → decrease τ).

[0041] In this embodiment, the five typical task classifications of the end-to-cloud routing are shown in Table 1.

[0042] Table 1

[0043] In step S4, the Dynamic Context Allocation Scheduler (DCAS) is used to allocate the budget using the following role × task two-dimensional token formula. B_alloc(u, θ) = B_total × [r(u) × g(θ)] / Σi[r(ui) × g(θi)] (3) Calculate the context token budget B_alloc(u, θ). In this embodiment, the context token budget B_alloc(u, θ) represents how much token budget each member and each task type can be allocated. In formula (3), B_total is the current available token budget (remaining space in the context window), u is the requesting member, θ is the current task type, r(u) is the role weight of member u, and g(θ) is the complexity multiplier of task θ. Member role weight r: family administrator r=1.2, ordinary member r=1.0, visitor r=0.6. Task complexity g: simple query g=0.7, multi-step instruction g=1.0, family decision g=1.3. Numerator = the current member's "priority" × "task complexity", denominator = the sum of the priorities of all concurrent members / tasks, and result = the proportion of the total budget that the current request can be allocated.

[0044] Reference Figure 4When performing step S5, semantic awareness context compression is performed: when the cumulative number of tokens exceeds the dynamic compression threshold C_threshold, the SACC compressor is triggered to calculate the ISM importance score according to formulas (4) and (5) and perform differential compression; health information (Health_Flag=true) is forced to skip compression.

[0045] In step S5, the semantically aware context compression algorithm SACC is used, based on the dialogue segment importance score (ISM) formula: I(sk) = w1·E(sk) + w2·A(sk) + w3·N(sk) + w4·P(sk, u) (4) Calculate the ISM importance score I(sk), where sk is the k-th dialogue segment, E represents sentiment intensity, A represents actionability, N represents entity novelty, and P represents personalization gain. The weighting coefficients w1, w2, w3, and w4 are w1=0.15, w2=0.35, w3=0.25, and w4=0.25, respectively. Actionability and novelty have higher weights, and segments containing decisions / facts are prioritized for retention.

[0046] Reference Figure 7 In step S5, the adaptive compression ratio formula is determined based on the task type: C_ratio(θ) = [B_avail B_mem B_cur] / B_hist · f(θ_task) (5) The task type-aware adaptive compression ratio C_ratio(θ) is calculated. The smaller the task type-aware adaptive compression ratio C_ratio(θ), the more aggressive the compression is in step S5. In formula (5), B_avail represents the total number of available context window tokens, B_mem represents the number of tokens occupied by memory injection, B_cur represents the number of real-time inputs in the current dialogue, B_hist represents the total number of original tokens in the history dialogue, and the numerator ([B_avail B_mem If B_cur] / B_hist)>1, it indicates sufficient space. The meaning of the task awareness factor f(θ_task) is as follows: Casual conversation θ1→f=0.3, aggressive compression is performed during step S5; Instruction control θ2→f=0.8, conservative compression is performed during step S5; Health query θ3→f=0.95, extremely conservative compression is performed during step S5; Information query θ4→f=0.6, moderate compression is performed during step S5. Paragraphs with Health_Flag=true force I_min_health=0.8, skipping compression during step S5.

[0047] In this embodiment, in step S6, inference is performed: the compressed context, recall memory, and current query are combined into a structured Prompt; based on the routing decision in S4, inference is performed on the edge LLM (simple task) or on the cloud large model (complex task) after privacy filtering.

[0048] In this embodiment, in step S7, result generation and dual writing to the end-cloud memory are performed: the reasoning result obtained in step S6 is acquired, the interactive content in the reasoning result is extracted, and the key information that meets the writing conditions (new facts / preference changes / family decisions) is synchronously written to the end-side memory cache (real-time) and the cloud CCM (asynchronous incremental synchronization, PIE desensitization processing is performed before writing, and the special default Health_Flag does not go to the cloud and requires authorization).

[0049] Specifically, during step S7, the Privacy Anonymization to Cloud Protocol (PIE) is executed. Specifically, before the memory entries are synchronized to the cloud CCM, the PIE module performs the following privacy processing flow: Named Entity Recognition (NER): Recognizes personal names in memorized text (→ replaces them with family role tags, such as "Dad / Mom / Xiaoming"), precise addresses (→ blurs them to the city level), and mobile phone numbers / ID numbers (→ fully hashed); Medical sensitive information: Memory entries with Health_Flag=true are not uploaded to the cloud by default and are only stored on the device; users can manually authorize specific health memories to be synchronized to the cloud (such as for cross-device allergy alerts). Encrypted transmission: Synchronized data is transmitted using AES-256 encryption, and cloud storage uses user key encryption (KMS management), which prevents cloud service providers from decrypting the original data; User data sovereignty: Users can delete all memories from CCM at any time. After executing "One-click memory clearing", the cloud data will be physically deleted within 24 hours.

[0050] In this embodiment, based on the execution of steps S1-S7, step S8 can also be executed to perform multimodal presentation of the middle screen: output the results in the form of voice broadcast + middle screen card + IoT linkage; update the terminal-side scene memory summary; trigger the cloud CCM incremental synchronization task as needed (if new memory is written).

[0051] Preferred embodiment Preferred Implementation Example 1: Multi-member Family Cloud Collaborative Daily Interaction Scenarios [Scenario Setting] A family of four (father, mother, and two children) has had the Family Smart Center Screen (FHS) installed and used for more than 30 days in the living room, and the system has completed basic preference learning; the master bedroom small screen and kitchen small screen are also bound under the same family account, and memory sharing is achieved through cloud CCM.

[0052] [Dialogue Sequence and Edge-to-Cloud Routing Decisions] Reference Figure 8 : T1 - Father: "Analyze our recent family spending and help me plan our family budget for next month." →CETR Assessment: C(q)=0.85 (Complex Financial Analysis), Pri(q)=0.7 (Including Family Financial Information), Net_score=0.9 (Good WiFi) → RouteScore_cloud=0.5×0.85 0.4 × 0.7 + 0.1 × 0.9 = 0.425 0.28 + 0.09 = 0.235 < τ_route = 0.3 → Due to the high privacy level, the route is executed via the edge LLM (including edge memory enhancement). If the parent authorizes this, the route can be routed to the cloud for enhanced analysis capabilities.

[0053] T2 - Mother: "I'm going to Shanghai on a business trip next week. Can you help me plan a 5-day, 4-night itinerary?" → CETR Assessment: C(q) = 0.9 (Complex Travel Planning), Pri(q) = 0.2 (General Geographic Information, Non-Household Sensitive Data), Net_score = 0.9 → RouteScore = 0.5 × 0.9 0.4 × 0.2 + 0.1 × 0.9 = 0.45 0.08 + 0.09 = 0.46 > 0.3 → Route to the cloud-based large model. The system retrieves preferences (business hotel / vegetarian preference) from the mother's personal memories, de-identifies them, and sends them to the cloud. The cloud-based large model's powerful planning capabilities are then used to generate a detailed itinerary.

[0054] T3 - Child: "I want to watch Paw Patrol!" → CETR Assessment: C(q)=0.05 (simple entertainment on-demand), Pri(q)=0.1 → RouteScore=0.075 (extremely low) → Can be handled by client-side LLM. SACC detects that the current token has reached 2800 / 4096, triggering compression (f(θ_entertainment)=0.3 aggressive compression), compressing the T1-T2 dialogue into a summary, freeing up a large amount of token space, and retaining the complete T3 instruction.

[0055] T4 - Memory Cross-Device Synchronization: When the mother returns to the bedroom, she asks on the bedroom screen, "What are we having for lunch today?" → The bedroom screen retrieves the shared family memories (including a summary of today's menu discussion) from the master bedroom screen and the living room screen via the cloud-based CCM, enabling seamless cross-device memory continuation without the mother needing to repeat the context.

[0056] [Key Parameter Calculation (CETR Route T2)] C(q)=0.9, Pri(q)=0.2, Net_score=0.9 →RouteScore=0.5×0.9 0.4×0.2+0.1×0.9=0.46>τ=0.3 → Cloud-based large model execution. Data sent is processed by PIE: Mother's name → "Family member B (female / administrator role)", home address → "a district in Shanghai".

[0057] Preferred Implementation Example 2: End-side Degradation Solution for Network Outage / Weak Network Scenarios [Scenario Setting] A family experiences a network outage (such as router failure or telecommunications maintenance), and the FHS device loses its cloud access.

[0058]

Execution Process

[0059] [Value] Compared to the pure cloud solution, the edge-cloud converged solution can still maintain basic personalized service capabilities in the event of network outage, thus solving the "network outage degradation" problem of Solution C.

[0060] Preferred Implementation Example 3: Home Healthcare Scenario (Extended Implementation Example) [Scenario Setting] A mother discovers that her child may be allergic to peanuts and consults the FHS for relevant information.

[0061] [Implementation Process] "Xiao Bao has been getting rashes every time he eats peanuts lately, which may be a peanut allergy. We shouldn't buy peanuts for our family anymore." The system detected the "peanuts → allergen" health entity chain → Health_Flag=true → CETR routing decision: Pri(q)=0.95 (highly sensitive medical information) → RouteScore is extremely low, forcing on-device processing → Health memories are synchronously written to the on-device personal memory layer (child's personal profile) and family shared memory layer (family list of prohibited ingredients) → Entries with Health_Flag=true are not uploaded to the cloud by default (unless the user actively authorizes it). All subsequent food recommendations will use allergy information for hard filtering.

[0062] In this embodiment, the following problem can be solved by executing steps S1-S7 or S1-S8: 1) Solve the problem of long-term memory confusion among multiple family members: a three-layer memory architecture (personal / family shared / scenario) + four-level storage on the edge and cloud, to achieve memory isolation and cross-device sharing among members; 2) Solving the information fidelity problem under the limited context window at the edge: SACC semantic-aware adaptive compression maximizes the context information density under a limited token budget; 3) Solve the problem of isolated memory between the device and the cloud: Device-side hot memory caching + cloud-based CCM cold memory backup + PIE privacy-de-identifying synchronization protocol to achieve persistent memory and cross-device sharing; 4) Solving the problem of missing task scheduling: The CETR module makes intelligent decisions on whether to execute LLM on the terminal side or large model in the cloud based on task complexity, privacy level and network status, taking into account both capacity limits and privacy protection.

[0063] In this embodiment, by executing steps S1-S7 or S1-S8, the following effects can be achieved: 1) The memory recall rate of multiple family members is significantly improved compared to the non-hierarchical memory solution, and the memory sharing capability across devices is greatly improved; 2) Semantic-aware compression significantly improves the effective information retention rate and greatly reduces the context token usage compared to traditional truncation schemes; 3) The response latency for simple tasks on the edge remains at the sub-second level, while the inference quality for complex tasks in the cloud is significantly better than that of the edge 7B model; 4) The PIE privacy desensitization mechanism ensures that sensitive family information is anonymized locally before synchronization, meeting the compliance requirements of the Personal Information Protection Law; 5) Memory data is persisted through cloud-based CCM, and memory is transferred without loss after device replacement / fault recovery, solving the data loss problem of pure end-side solutions.

[0064] Specifically: (1) Through the three-layer long memory architecture perceived by family members and the end-to-cloud collaborative retrieval, that is, the family interactive memory is divided into three layers (PLM_P / PLM_S / EPM) according to the two dimensions of "attribution + timeliness". The hot cache is retained on the end side, and the historical cold data is archived in the cloud CCM after PIE desensitization. The end-to-cloud collaborative cross-layer semantic retrieval is performed by formula (1) five-dimensional comprehensive score (semantic similarity, member relevance, time decay, access frequency, entity novelty). It supports the two-level strategy of fast retrieval on the end side + supplementary retrieval in the cloud, which can significantly improve the recall rate of multi-member memory, achieve lossless migration of memory across devices, and complete recovery of memory after device replacement / failure.

[0065] (2) Through the context allocation scheduler (DCAS) that dynamically adjusts according to family roles and task types, namely the joint product allocation model based on member role weight r(u) and task complexity function g(θ) (Formula 2), the token budget is dynamically tilted towards family administrators and high-complexity tasks. When multiple family members interact simultaneously, context resources are tilted towards important family decisions, ensuring that information utility is maximized under a limited token window, significantly improving context utilization, significantly improving the integrity of information for family decision-making tasks, and greatly increasing token resource utilization.

[0066] (3) The edge-cloud task router (CETR) based on complexity-privacy-network three dimensions is used. That is, the task complexity C(q), privacy level Pri(q) and network quality Net_score are comprehensively evaluated by formula (3), and the edge-cloud execution comprehensive score RouteScore_cloud(q) is calculated. The dynamic decision is made to execute the current task on the edge lightweight LLM or the cloud large model. The routing strategy takes into account both the upper limit of inference capability (highly complex tasks go to the cloud) and privacy protection (high privacy tasks stay on the edge); when the network is disconnected, it automatically degrades to the edge for execution, and automatically performs incremental synchronization after the network is restored. This achieves that the inference quality of high complex tasks is significantly better than that of the pure edge 7B model, and the high privacy tasks retain edge processing to ensure data security. Basic personalized services can still be provided in the scenario of network disconnection.

[0067] (4) Through semantically aware adaptive context compression (SACC), namely through two innovative sub-modules: (A) an ISM importance scoring model based on four factors (emotional intensity / actionality / entity novelty / personality gain) (Formula 4); (B) an adaptive compression rate control factor f(θ_task) based on task type awareness (Formula 5). Low-scoring redundant segments are compressed first, while high-scoring key segments are retained; segments with Health_Flag=true are forced to have a minimum retention threshold, and are never compressed or lost, achieving a significantly higher effective information retention rate compared to FIFO truncation, a substantial reduction in context token occupancy, and the Health_Flag mechanism to achieve zero compression loss of key health information.

[0068] (5) The privacy-graded desensitization cloud synchronization protocol (PIE+CCM) is used, which introduces the privacy information extractor (PIE). Before the memory is synchronized to the cloud CCM, NER entity recognition + graded desensitization is performed (personal name is replaced with role label, address is obfuscated, medical information is retained locally or uploaded with encryption after user authorization); memory with Health_Flag=true is not uploaded to the cloud by default; AES-256 encryption is used throughout the transmission; users can delete all memories in the cloud with one click (physical deletion in 24 hours), realizing zero cloud upload of sensitive health information to local storage; other memories can be safely backed up after desensitization; cloud data can be completely deleted, and users have control over their data sovereignty.

[0069] A computer program can be written to execute the member-aware home smart screen device context processing method of this embodiment, and then written into a computer device or storage medium. When the computer program is read and run, the member-aware home smart screen device context processing method of this embodiment is executed, thereby achieving the same technical effect as the member-aware home smart screen device context processing method of the embodiment. For example, when the computer program is written into the memory of the computer device, the processor can read the computer program from the memory and execute the member-aware home smart screen device context processing method of this embodiment.

[0070] It should be noted that, unless otherwise specified, when a feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. Furthermore, the descriptions of "upper," "lower," "left," and "right" used in this disclosure are only relative to the relative positional relationships of the components of this disclosure in the accompanying drawings. The singular forms "a" and "the" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. Moreover, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this embodiment specification is only for describing particular embodiments and is not intended to limit the invention. The term "and / or" as used in this embodiment includes any combination of one or more of the associated listed items.

[0071] It should be understood that although the terms first, second, third, etc., may be used to describe various elements in this disclosure, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, a first element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as a first element. The use of any and all instances or exemplary language (“e.g.,” “such as,” etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and, unless otherwise required, does not impose a limitation on the scope of the invention.

[0072] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0073] Furthermore, the procedures described in this embodiment can be performed in any suitable order unless otherwise indicated by this embodiment or clearly contradicted by the context. The procedures (or variations and / or combinations thereof) described in this embodiment can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. A computer program includes multiple instructions executable by one or more processors.

[0074] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices, etc. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention of this embodiment includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques of the invention, the invention also includes the computer itself.

[0075] A computer program can be applied to input data to perform the functions of this embodiment, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects generated on the display.

[0076] The above are merely preferred embodiments of the present invention. The present invention is not limited to the above-described embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, as long as they achieve the technical effects of the present invention by the same means, should be included within the scope of protection of the present invention. Within the scope of protection of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.

Claims

1. A context processing method for a home smart central screen device based on member perception, characterized in that, The context processing method for the home smart central screen device based on member awareness includes: Identify the currently interacting members of the smart home central display device; Collect the current scene context of the smart home central screen device, and construct an initial context vector based on the current scene context; Based on the initial context vector, a search is performed on the edge of the smart home screen device. When the hit rate of the edge memory determined by the search is lower than the hit rate threshold, a supplementary search is asynchronously initiated to the cloud. The search results of the edge and / or the cloud are comprehensively scored, and the Top-K memory is recalled and injected into the context based on the scoring results. The complexity and privacy classification of the current task of the smart home screen device are evaluated, the edge-cloud execution score is calculated, and it is determined whether the inference is performed on the edge or the cloud. A context token budget is allocated to the executor, where the executor is the edge or the cloud. When the cumulative number of tokens for the current task of the smart home screen device exceeds the dynamic compression threshold, the ISM importance score is calculated, and differential compression is performed on the context of the current task to obtain the compressed context. The compressed context, recalled memory, and current query are combined into structured prompts, and inference is performed at the executor. Obtain the reasoning result, perform memory extraction on the interactive content in the reasoning result, obtain key information that meets the writing conditions, perform desensitization processing on the key information, and synchronously write the key information to the terminal and the cloud.

2. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The process of obtaining the lyrics text to be translated also includes: The key information is output in a multimodal linkage manner to update the context memory summary on the device side; and the incremental synchronization task in the cloud is triggered as needed.

3. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The currently interacting members of the identified smart home central screen device include: The currently interacting member is identified through voiceprint, person, or device binding; Initiate the memory space for the currently interacting member and determine whether the currently interacting member is being used for the first time. If so, synchronize the historical memory of the currently interacting member.

4. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The comprehensive scoring of the search results from the endpoint and / or the cloud includes: According to the formula Score(m,q,u) =α·Sim(Em, Eq)+β·Rel(m,u)·e^(-λΔt)+γ·Freq(m)+δ·Novelty(m,q) The comprehensive score Score(m,q,u) is calculated; where m is memory, q is query, u is user, Em is memory semantics, Eq is query semantics, Sim is cosine similarity; Rel(m,u) is member relevance weight; e^(-λΔt) is time decay; Freq(m) is access frequency; Novelty(m,q) is information novelty; α, β, γ, δ are weight coefficients, satisfying α+β+γ+δ=1.

5. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The process of assessing the complexity and privacy of the current task of the smart home screen device, calculating the edge-cloud execution score, and determining whether the inference is performed on the edge or in the cloud includes: According to the formula RouteScore_cloud(q)=w_c·C(q)-w_p·Pri(q)+w_n·Net_score The decision score RouteScore_cloud(q) is calculated, where C(q) is the task complexity score, Pri(q) is the privacy level, Net_score is the overall score of the current network quality, and w_c, w_p, and w_n are weight coefficients. Set the routing threshold τ_route; If the decision score RouteScore_cloud(q) is greater than the routing threshold τ_route, the inference will be executed in the cloud; otherwise, the inference will be executed on the device.

6. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The allocation of the context token budget to the executor includes: According to the formula B_alloc(u, θ) = B_total × [r(u) × g(θ)] / Σi[r(ui) × g(θi)] The context token budget B_alloc(u, θ) is calculated; where B_total is the total available token budget, u is the requesting member, θ is the current task type, r(u) is the role weight of member u, g(θ) is the complexity multiplier of task θ, r(ui) is the role weight of the i-th member u, and g(θi) is the complexity multiplier of the i-th task θ.

7. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The calculation of the ISM importance score includes: According to the formula I(sk)=w1·E(sk)+w2·A(sk)+w3·N(sk)+w4·P(sk, u) The ISM importance score I(sk) is calculated, where sk is the k-th dialogue segment, E represents emotional intensity, A represents actionability, N represents entity novelty, P represents personalization gain, and w1, w2, w3, and w4 are weights.

8. The context processing method for a home smart central screen device based on member perception according to claim 1, characterized in that, The differential compression of the context of the current task includes: According to the formula C_ratio(θ) = [B_avail B_mem B_cur] / B_hist · f(θ_task) The task type-aware adaptive compression ratio C_ratio(θ) is calculated; where B_avai is the total number of available context window tokens, B_mem is the number of tokens occupied by memory injection, B_cur is the real-time input of the current dialogue, B_hist is the total number of original tokens in the history dialogue, and f(θ_task) is the task-aware factor. Based on the task type-aware adaptive compression ratio C_ratio(θ), compression is performed on the context of the current task.

9. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store at least one program, and the processor being used to load at least one program to execute the context processing method for a home smart central screen device based on member awareness as described in any one of claims 1-8.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the member-aware home smart central screen device context processing method according to any one of claims 1-8.