A cross-hardware AI memory interaction system

Abstract

The application belongs to the field of AI terminal hardware, and provides a cross-hardware AI memory interaction system, which comprises a core module, the core module comprises a main control calculation unit, a magnetic attraction array and a PogoPin physical connection interface, a wireless communication module, an NFC reading and writing module, an audio interaction unit, an encrypted storage unit and a power management unit, the main control calculation unit comprises an ARM architecture processor and a hardware neural network processing unit, and is used for lightweight AI inference, memory data management and Agent image scheduling, and the magnetic attraction array and the PogoPin physical connection interface are used for being adsorbed on the back of a mobile terminal; by the system, user lifelong memory and Agent personality are all encrypted and stored in a hardware that can be carried, the user has complete ownership and control right, can freely migrate across devices, and does not depend on any manufacturer cloud; in cooperation with biological feature encryption, the risk of privacy leakage is eliminated from the physical layer.

Description

Technical Field

[0001] This invention relates to AI terminal hardware technology, specifically to a cross-hardware AI memory interaction system. Background Technology

[0002] With the development of edge AI and large-scale model technology, portable AI hardware has gradually become a hot topic in the industry. Existing products are mainly divided into two architectures:

[0003] The first type is a pure cloud computing architecture, which delegates all speech recognition and semantic reasoning to remote servers. This solution is heavily reliant on the network; the function fails when the network is disconnected. Furthermore, all of the user's private data, including voice, behavior, and memories, is uploaded to a third-party platform, posing a high risk of data leakage and misuse, and users do not have data sovereignty.

[0004] The second type features a built-in independent processor architecture, where the device integrates a complete computing chip. However, due to limitations in portability and battery capacity, it is difficult to balance processor performance, storage space, and battery life. This results in high manufacturing costs, low computing power limits, and an inability to support users' lifetime memory storage and complex reasoning tasks.

[0005] When using existing Pogopin magnetic AI memory cross-hardware interaction systems, users' AI memories and personality data are either fixed locally on a single device or stored in the manufacturer's cloud, making it impossible to independently possess or freely migrate across devices. Changing devices or platforms results in the loss of all historical interaction data, preventing users from truly owning their own digital memory assets. The personal AI terminal and home smart devices operate independently, with data not being shared and memories not being synchronized. Users need to use multiple independent systems in different scenarios, resulting in a fragmented experience and an inability to form a unified personal digital identity. Summary of the Invention

[0006] The purpose of this invention is to provide a cross-hardware AI memory interaction system to solve the problems of high privacy risks in existing AI hardware and the lack of user sovereignty over memory data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a cross-hardware AI memory interaction system, comprising a core module, wherein the core module includes:

[0008] The main control computing unit includes an ARM architecture processor and a hardware neural network processing unit, used for lightweight AI inference, memory data management and agent image scheduling;

[0009] The terminal connection interface is used for bidirectional encrypted synchronization of memory data and agent image with the parasitic agent software in the mobile terminal, and to call the neural network processing unit of the mobile terminal to assist in the execution of AI inference tasks.

[0010] The encrypted storage unit uses encryption algorithms to store a user-exclusive Agent image library and multi-level lifelong memory data. The decryption key for the memory data is derived from the user's biometrics and the device's hardware fingerprint.

[0011] The core module is an independent physical carrier of the user's lifelong memory and Agent personality, which can be plugged in and moved between mobile terminals and home global computing power hosts;

[0012] The terminal only temporarily caches the minimum memory fragment required for the current task, and automatically clears it after the task ends. Core memory data does not reside locally on the terminal.

[0013] Furthermore, the PogoPin interface includes:

[0014] Pin 1: VBUS power pin, supports bidirectional 5V power transmission;

[0015] Pin 2: GND grounding pin;

[0016] Pin 3: USB 2.0 D+ data pin;

[0017] Pin 4: USB 2.0 D - Data Pin;

[0018] Pins 3 and 4 form a high-speed data channel for encrypted transmission of memory data, Agent configuration files, and AI inference task flows.

[0019] Pin 5: Custom serial communication TX pin, used to output sensor data and control commands;

[0020] Pin 6: Custom serial communication RX pin, used to receive computing power scheduling instructions and inference results from the mobile terminal.

[0021] Furthermore, the hardware is adapted to a three-tier computing power collaborative architecture:

[0022] Portable lightweight computing layer: It runs entirely on the built-in neural network processing unit of the core module and can run basic interactive functions offline;

[0023] Mobile auxiliary computing power layer: When attached to a mobile terminal, the software agent calls on the mobile terminal's computing power, while the magnetically attached hardware only stores the contextual decision memory and executes medium-complexity reasoning tasks;

[0024] Home-based heavy computing layer: When connected to a home global computing host, complex reasoning, long memory association, and multi-device scheduling tasks are offloaded to the host for execution, and the core module is only responsible for memory storage and interactive input and output.

[0025] Furthermore, the three-tiered computing power collaborative architecture also includes a cloud-enhanced subsystem and a knowledge localization and accumulation mechanism:

[0026] When connected to the internet, the API of a third-party cloud-based large model is called to process the de-identified common questions. The returned common knowledge fragments are extracted, vectorized, and written into a local encrypted storage unit, realizing knowledge accumulation that can be used permanently offline with a single call. The original privacy memory data is never fully uploaded to the public cloud.

[0027] Furthermore, the memory system in the encrypted storage unit adopts a three-layer architecture and supports cross-device encrypted synchronization:

[0028] Working memory layer: Stores the current session context and is automatically cleared when the session ends;

[0029] Contextual memory layer: Stores historical dialogues, event records, and emotion tags in the form of a vector database, supporting semantic retrieval;

[0030] Semantic memory layer: Stores user preferences, behavioral habits, and thinking patterns in the form of structured data;

[0031] The memory data is permanently stored in the core module and can be synchronized to the mobile parasitic agent and the home global computing power host. The synchronization process is encrypted throughout.

[0032] Furthermore, the memory system also includes:

[0033] Time decay mechanism: Memory weights are calculated using an exponential decay function, and low-weight memories are automatically archived and compressed;

[0034] Memory fusion mechanism: Periodically merge fragmented memories on the same topic into summary memories;

[0035] Continuous learning mechanism: continuously update user features in the semantic memory layer based on user historical interactions.

[0036] Furthermore, it also includes building a dual-track agent ecosystem, including a digital software agent subsystem and a physical NFC shell agent subsystem;

[0037] The digital software Agent subsystem includes an online Agent application store and a local Agent image library. Users can access the online store through the Agent software on their mobile terminals to browse, purchase, and download digital Agents with different functions. The downloaded Agent images are stored in the encrypted storage unit of the core module and support one-click switching of the software interface.

[0038] The physical NFC shell Agent subsystem includes a replaceable NFC Agent shell, with a passive NFC tag embedded inside the shell, storing a unique authorized identifier for the Agent. After the core module reads the tag, it automatically loads the complete functional image of the corresponding Agent.

[0039] Furthermore, the Agent configuration set stored in the NFC tag includes:

[0040] Agent Function Module ID: Points to the corresponding function plugin in the core module or home host, such as office assistant, life manager, sports coach, learning tutor, etc.

[0041] Personality parameter set: Defines the tone of voice, word preferences, response rhythm, and logical tendencies of the agent's interaction;

[0042] Voice packet identifier: points to the corresponding TTS speech synthesis model in the core module's storage unit;

[0043] Permission configuration: Defines the memory levels and functional permissions that this Agent can access;

[0044] The core module completes a full Agent switch within 1 second after reading the NFC tag. After the switch, the interaction logic, voice style, and available functions take effect simultaneously. Agents with different functions can be developed through the open SDK and distributed by binding customized shells.

[0045] Furthermore, it also includes a home host docking station, which is a fixed terminal for home use and includes:

[0046] Speaker arrays and circular microphone arrays are used for far-field voice interaction;

[0047] High-speed synchronization interface: When the core module is placed on the base, an encrypted channel is automatically established to fully synchronize the memory data and the full Agent image library with the home global computing power host;

[0048] Computing offload channel: forwards complex inference and multi-device scheduling tasks to home global computing power hosts for execution;

[0049] Charging unit and independent auxiliary computing unit;

[0050] Independent auxiliary computing unit: When the core module is not present, it can call upon the computing power of a home host to provide basic AI interaction services.

[0051] Compared with existing technologies, the present invention provides a cross-hardware AI memory interaction system in which the user's lifelong memory and Agent personality are fully encrypted and stored in portable hardware. The user has complete ownership and control, and can freely migrate across devices without relying on any manufacturer's cloud. Combined with biometric encryption, the risk of privacy leakage is eliminated at the physical level.

[0052] Through three-level computing power scheduling, the device leverages the idle NPU of the mobile device to enhance computing power in the on-the-go state, eliminating the need for its own high-performance chip and significantly reducing device size, power consumption, and cost. In the home state, it connects to the home host to unleash powerful computing power, balancing portability and performance limits. The digital software library solves the pain point of multi-agent portability: there is no need to carry multiple physical shells when going out, and all purchased agents can be switched with one click on the mobile device, which is highly scalable and has a fast update cycle. The physical NFC shell provides unique value: one-touch blind switching, offline physical distribution, and premium pricing for collectible co-branded items, forming a complete business loop that links online and offline. Analogous to the ecosystem model of "digital APP + physical peripherals", it has richer monetization paths. The on-the-go and home scene memory is automatically synchronized and the agents are seamlessly migrated. Users have a unified personal digital identity, which solves the industry pain point of fragmented multi-device scenarios and provides a unified user personality and memory entry point for whole-house smart home. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0054] Figure 1 This is a block diagram of the internal functional structure of the core module in this invention;

[0055] Figure 2 This is a flowchart illustrating the working principle of NFC-Agent shell switching in this invention;

[0056] Figure 3 This is a simplified schematic diagram of the three-level computing power collaborative architecture in this invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0058] As attached Figure 1 To be continued Figure 3 As shown:

[0059] Example 1:

[0060] This invention provides a cross-hardware AI memory interaction system, including a core module, the core module comprising:

[0061] The main control computing unit includes an ARM architecture processor and a hardware neural network processing unit, used for lightweight AI inference, memory data management and agent image scheduling;

[0062] The terminal connection interface is used for bidirectional encrypted synchronization of memory data and agent image with the parasitic agent software in the mobile terminal, and to call the neural network processing unit of the mobile terminal to assist in the execution of AI inference tasks.

[0063] The encrypted storage unit uses encryption algorithms to store a user-exclusive Agent image library and multi-level lifelong memory data. The decryption key for the memory data is derived from the user's biometrics and the device's hardware fingerprint.

[0064] The core module is an independent physical carrier of the user's lifelong memory and Agent personality, which can be plugged in and moved between mobile terminals and home global computing power hosts;

[0065] The terminal only temporarily caches the minimum memory fragment required by the current task, and automatically clears it after the task ends. Core memory data does not reside locally on the terminal.

[0066] The magnetic array and PogoPin physical connection interface are used to attach to the back of the mobile terminal. The PogoPin interface contains no fewer than 6 contacts for bidirectional encrypted synchronization of memory data and agent image with the parasitic agent software within the mobile terminal, and for calling the mobile terminal's neural network processing unit to assist in executing AI inference tasks. Specifically, it is defined as follows:

[0067] Pin 1 VBUS: Bidirectional 5V power transmission, which can draw power from the mobile device or provide emergency reverse power to the mobile device.

[0068] Pin 2 GND: Common ground;

[0069] Pin 3 (D+) and Pin 4 (D-): USB 2.0 high-speed data channel with a transmission rate of ≥480Mbps, used for encrypted transmission of memory data, Agent configuration files, and inference task streams;

[0070] Pin 5 (TX): Custom serial transmission channel for transmitting sensor data and control commands;

[0071] Pin 6 RX: Custom serial receive channel, receiving mobile terminal scheduling instructions and inference results;

[0072] After attaching to the mobile device, it automatically establishes a connection with the parasitic agent software pre-installed on the mobile device to achieve bidirectional encrypted data synchronization and collaborative computing power utilization.

[0073] The wireless communication module supports Wi-Fi and Bluetooth communication protocols for data interaction with home global computing power hosts and Bluetooth audio devices, as well as for downloading digital agent images from online agent application stores. It integrates a Wi-Fi 5 dual-band chip and a Bluetooth 5.0 chip. Wi-Fi is used to connect to the home global computing power host, access cloud-based de-identified APIs, and download agent images; Bluetooth is used to connect to peripherals such as Bluetooth headsets and wristbands, and supports A2DP / HFP audio protocols.

[0074] The audio interaction unit includes a dual-microphone array (supporting noise reduction and sound pickup), a bone conduction vibration transducer, and a drive circuit. The bone conduction unit is attached to the back cover of the mobile device through a silicone coupling layer, and the vibration is transmitted to the user's hand through the mobile device shell, enabling private calls without the need for an in-ear device. It can also switch to Bluetooth mode to output audio.

[0075] The power management unit has a built-in thin and light lithium battery, supports Qi2 wireless charging, PogoPin bidirectional power supply, and dynamic power consumption adjustment.

[0076] The encrypted storage unit stores a user-exclusive Agent image library and multi-level lifetime memory data using encryption algorithms. The decryption key for the memory data is derived from the user's biometric features and the device's hardware fingerprint, and can only be unlocked and accessed by the user. It is equipped with a large-capacity flash memory chip and uses AES-256 full disk encryption. The decryption key is derived from the user's voiceprint features and the device's hardware fingerprint. Each access to the memory must be verified by voiceprint. If the device is lost, third parties cannot crack and read it. The stored content includes: a three-layer lifetime memory database, a local Agent image library, a TTS voice model package, and system firmware.

[0077] The core module is an independent physical carrier of the user's lifelong memory and Agent personality. It can be plugged and transferred between mobile terminals and home global computing power hosts. It supports dual-track switching between digital software Agent and physical NFC shell Agent, realizing seamless connection between memory data and Agent functions in on-the-go and home scenarios.

[0078] Preferred hardware-adapted three-tier computing power collaborative architecture:

[0079] Portable lightweight computing layer: It runs entirely on the built-in neural network processing unit of the core module and can run basic interactive functions offline;

[0080] Mobile-assisted computing layer: When attached to a mobile terminal, the software agent utilizes the mobile terminal's computing power, while the magnetically attached hardware only stores contextual decision-making memories and executes medium-complexity inference tasks. When the core module is attached to the back of the mobile terminal, the system automatically enters the mobile-assisted computing layer working mode. Through a high-speed encrypted USB channel established via the PogoPin interface, the core module and the mobile terminal's parasitic agent software collaborate: the core module is only responsible for storing the user's contextual and semantic memories and the agent image, and desensitizing input commands, ensuring privacy data never leaves the hardware; the mobile terminal, on the other hand, utilizes its NPU to execute medium-complexity inference tasks (such as information summarization, simple planning, and enhanced retrieval generation). After the inference results are encrypted and sent back to the core module, the mobile terminal immediately clears all temporary caches, ensuring a traceless process. This mode fully utilizes the mobile terminal's idle computing power, keeping the core module low-power and small in size, while balancing inference performance and data sovereignty security in portable scenarios, achieving a balance between performance, privacy, and portability.

[0081] Home-based heavy computing layer: When connected to a home global computing host, complex reasoning, long memory association, and multi-device scheduling tasks are offloaded to the host for execution, and the core module is only responsible for memory storage and interactive input and output.

[0082] Preferably, the three-tier computing power collaborative architecture also includes a cloud-based enhancement subsystem and a knowledge localization and accumulation mechanism:

[0083] When connected to the internet, the API of a third-party cloud-based large model is called to process the de-identified common questions. The returned common knowledge fragments are extracted, vectorized, and written into a local encrypted storage unit, realizing knowledge accumulation that can be used permanently offline with a single call. The original privacy memory data is never fully uploaded to the public cloud.

[0084] Employing an edge-cloud collaborative logic, it integrates two AI systems: one utilizes the mobile device's NPU to process personal emotions, and the other integrates with other large-scale models, such as:

[0085] 1. You asked a complex question (e.g., analyze the impact of oil prices on the US stock market in 2026).

[0086] 2. The NPU determines that it has "insufficient computing power" and sends the encrypted request to DeepSeek in the cloud;

[0087] 3. DeepSeek provided an excellent answer;

[0088] 4. Key Action: At the same time as this response is generated, the system will automatically extract the "semantic essence" of the conversation, send it back and store it in the individual's database.

[0089] Specifically, the memory system in the encrypted storage unit adopts a three-tier architecture and supports cross-device encrypted synchronization:

[0090] Working memory layer: Stores the current session context and is automatically cleared when the session ends;

[0091] Contextual memory layer: Stores historical dialogues, event records, and emotion tags in the form of a vector database, supporting semantic retrieval;

[0092] Semantic memory layer: Stores user preferences, behavioral habits, and thinking patterns in the form of structured data;

[0093] The memory data is permanently stored in the core module and can be synchronized to the mobile parasitic agent and the home global computing power host. The synchronization process is encrypted throughout.

[0094] When connected to the network, if the system determines that the current task exceeds the processing capacity of local and mobile auxiliary computing power (such as complex knowledge question answering, deep analysis, etc.), it automatically triggers the end-to-cloud collaborative working mode. The core module first performs local privacy desensitization processing on user input, removing all personally identifiable information such as name, location, and contact information through sensitive word filtering and semantic recognition models. Only the desensitized general questions are used to call the third-party cloud large model API through an encrypted channel to obtain the reasoning results. At the same time, the system automatically extracts general knowledge fragments from the returned results, vectorizes them, and writes them into the local encrypted storage unit to achieve local knowledge accumulation. The entire process strictly follows the principle of "original privacy memory data is never fully uploaded to the public cloud." The cloud only processes abstract semantic requests without identity references, and after each call, local knowledge assets that can be reused offline are accumulated. Subsequent similar questions can be directly retrieved from the local vector library without repeated network calls, thus achieving the gradual accumulation of knowledge with "single network call, permanent offline availability." Under the premise of ensuring absolute privacy and security, it significantly reduces cloud dependence and continuously enhances the local intelligence level.

[0095] Working principle: After the user's voice commands are captured by the microphone and processed for noise reduction, the main control unit first performs a complexity classification:

[0096] Simple commands (schedule reminders, memory retrieval, daily conversations): The core module's local NPU directly runs a ≤3B parameter quantization model to complete inference, and the results are output through bone conduction. The entire process is offline, and the data does not leave the hardware. Interaction records are encrypted and written to the local memory system.

[0097] Medium-complexity instructions (information organization, simple planning): sent to the mobile parasitic agent via PogoPin, calling the mobile NPU to run the 3B-7B parameter model to complete inference, and the results are sent back to the core module output;

[0098] For general knowledge questions that exceed local computing power: In the network state, privacy desensitization is performed first. After removing all personal sensitive information through local sensitive word filtering and semantic recognition model, the results are obtained by calling the third-party cloud big model API. At the same time, general knowledge fragments in the results are automatically extracted, vectorized and written into the local memory bank. Subsequent similar questions can be directly retrieved locally without repeated network access.

[0099] All memory data generated by the interaction is encrypted and stored locally in the core module in real time, and periodically synchronized to the home host for backup.

[0100] Example 2:

[0101] This embodiment is basically the same as the previous embodiment, except that it also includes the construction of a dual-track agent ecosystem, including a digital software agent subsystem and a physical NFC shell agent subsystem. The digital agent and the physical shell agent are similar to Nintendo's digital game store and physical game cartridges.

[0102] The digital software agent subsystem includes an online agent app store and a local agent image library. Users can access the online store through the agent software on their mobile devices to browse, purchase, and download digital agents with different functions. Downloaded agent images are stored in the encrypted storage unit of the core module and support one-click switching between software interfaces. The digital software agent is the primary form of daily use, relying on the mobile agent software and the online agent app store.

[0103] Online Agent App Store: The official cloud service platform showcases all categories of agents, including scenarios such as office assistant, life manager, sports coach, learning tutor, psychological counselor, and game companion; third-party developers can develop agents through the open SDK, which can be listed on the store after review and monetization using a revenue-sharing model.

[0104] Local Agent Image Library: Digital agents purchased by users can be downloaded to the encrypted storage unit of the core module via Wi-Fi. Frequently used agents are pre-stored locally and can be switched for use offline. The home global computing power host stores the full agent image library. The home computing power host serves not only consumers but also businesses. It stores environmental decision data of robots entering the factory, which is automatically synchronized with the robot's onboard hardware.

[0105] One-click software switching: Users can select the target agent from the local image library through the visual interface of the parasitic agent on the mobile terminal or by voice command, and complete the full switch within 1 second - including functional logic, interactive personality, voice tone, and memory access permissions all take effect simultaneously.

[0106] In out-of-home scenarios, users do not need to carry any physical accessories. They can switch freely between multiple agents using only mobile software, completely solving the pain point of the inconvenience of carrying multiple agents.

[0107] The physical NFC shell Agent subsystem includes a replaceable NFC Agent shell. The shell has an embedded passive NFC tag that stores the Agent's unique authorization identifier. After the core module reads the tag, it automatically loads the complete functional image of the corresponding Agent. The replaceable NFC shell serves as a supplementary form and is magnetically attached to the outside of the core module. Different shells with different appearances and themes correspond to Agents with different functions, enabling "switching with a single touch".

[0108] Shell structure: The shell is a magnetic decorative protective shell, and the material can be customized to metal, resin, fabric, etc. It has an embedded passive NFC tag (compliant with ISO14443A standard). The tag has a built-in Agent unique authorized ID and basic configuration set. It does not require power supply and can be read when it is close to the core module.

[0109] Agent switching principle: When the user changes the shell, the NFC reading and writing module of the core module automatically reads the Agent identifier in the tag, pulls the complete image of the corresponding Agent from local storage or home host, and completes the full switching of functions, personality and voice within 1 second.

[0110] Typical scenario: Business black shell → Office assistant Agent: Proficient in schedule management, email processing, meeting minutes, and speaks in a professional and rigorous tone;

[0111] Vibrant orange shell → Sports Coach Agent: Responsible for exercise plans, diet records, and action reminders; speaks in a positive and energetic tone.

[0112] Soft pink casing → Life Butler Agent: Manages household chores, family matters, and shopping lists, with a friendly tone.

[0113] Preferably, the dual-track agent ecosystem supports physical-digital authorization interoperability:

[0114] After purchasing a physical NFC case, users can bind their accounts by reading the tag through the core module, which automatically unlocks the permanent use of the corresponding digital agent. It can also be used via software without the case.

[0115] After purchasing the digital Agent, users can also purchase the corresponding physical casing separately as a physical accessory and a quick switching platform. Both are authorized from the same source and their data is interoperable.

[0116] Specifically, the Agent's feature configuration set includes:

[0117] Agent functional module ID, personality parameter set, voice pack identifier, memory access permission configuration, and functional plugin list;

[0118] When switching agents in core modules, the entire process of changing functional logic, interaction characteristics, voice style, and permission scope is quickly and completely implemented.

[0119] It supports third-party developers to develop agents for different scenarios through the open SDK, upload them to online app stores for review and distribution, or bind them to customized NFC shells for offline physical distribution.

[0120] Working principle: In out-of-home scenarios, users do not need to carry any physical accessories. They can freely switch between multiple agents using only the mobile software, completely solving the pain point of inconvenience in carrying multiple agents. When the user places the case close to the core module or attaches it, the NFC reader automatically reads the agent identifier in the tag. After verification and authorization, it loads the full configuration of the corresponding agent from the local image library and takes effect within 1 second. Typical scenarios: Place the business case on the office desk, and when you are at work, simply place the hardware on it to automatically switch to work mode; carry the sports case when exercising, and switch with a single tap while running, without having to take out the mobile device to unlock it, achieving blind switching.

[0121] Example 3:

[0122] This embodiment is basically the same as the previous embodiment, except that it also includes a home host docking dock. The dock is a fixed terminal for home scenarios and includes:

[0123] Speaker arrays and circular microphone arrays are used for far-field voice interaction;

[0124] High-speed synchronization interface: When the core module is placed on the base, an encrypted channel is automatically established to fully synchronize the memory data and the full Agent image library with the home global computing power host;

[0125] Computing offload channel: forwards complex inference and multi-device scheduling tasks to home global computing power hosts for execution;

[0126] Charging unit and independent auxiliary computing unit.

[0127] Independent auxiliary computing unit: When the core module is not present, it can call upon the computing power of a home host to provide basic AI interaction services.

[0128] Working principle: This embodiment is a fixed terminal for home use, placed on an indoor desktop, serving as a connection hub between portable hardware and the home global computing power host. The base's functional components are as follows:

[0129] 1. High-speed synchronization and charging area: After the core module is placed, a high-speed encrypted connection is established through the contact points, and the newly generated memory is automatically synchronized to the home global computing host, while the hardware is fast charged (45W).

[0130] 2. Far-field interaction unit: Built-in 6-microphone ring array + full-range speaker array, supports 360° far-field voice wake-up and external speaker interaction. When the core module is not present, it can call on the computing power of the home host to provide public AI services for the home.

[0131] 3. Computing power offloading channel: After connecting to the home global computing power host, heavy tasks such as multi-robot scheduling, whole-house space planning, and long memory deep analysis can be forwarded to the host for execution. The core module and the base are only responsible for interactive input and output, fully releasing the advantages of strong computing power at home;

[0132] 4. Extended sensing unit: integrates millimeter-wave human presence radar and environmental sensors to achieve seamless interaction by automatically waking up when a person arrives and automatically going into sleep mode when a person leaves;

[0133] Through this embodiment, the core module can achieve a seamless switch between "carrying it with you when you go out and placing it on the base when you come home": in the carry-on state, it relies on its own light computing power + mobile terminal auxiliary computing power, and in the home state, it connects to the home host to obtain full strong computing power, and the memory data is automatically synchronized throughout the process, so that the user can complete the scene transition without feeling.

[0134] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A cross-hardware AI memory interaction system, characterized in that, Includes a core module, which includes: The main control computing unit includes an ARM architecture processor and a hardware neural network processing unit, used for lightweight AI inference, memory data management and agent image scheduling; The terminal connection interface is used for bidirectional encrypted synchronization of memory data and agent image with the parasitic agent software in the mobile terminal, and to call the neural network processing unit of the mobile terminal to assist in the execution of AI inference tasks. The encrypted storage unit uses encryption algorithms to store a user-exclusive Agent image library and multi-level lifelong memory data. The decryption key for the memory data is derived from the user's biometrics and the device's hardware fingerprint. The core module is an independent physical carrier of the user's lifelong memory and Agent personality, which can be plugged in and moved between mobile terminals and home global computing power hosts; The terminal only temporarily caches the minimum memory fragment required for the current task, and automatically clears it after the task ends. Core memory data does not reside locally on the terminal.

2. The cross-hardware AI memory interaction system according to claim 1, characterized in that, The terminal connection interface is a magnetic array and a PogoPin physical connection interface, used to attach to the back of the mobile terminal. The PogoPin interface includes no fewer than 6 contact points. The PogoPin interface includes: Pin 1: VBUS power pin, supports bidirectional 5V power transmission; Pin 2: GND grounding pin; Pin 3: USB 2.0 D+ data pin; Pin 4: USB 2.0 D - Data Pin; Pins 3 and 4 form a high-speed data channel for encrypted transmission of memory data, Agent configuration files, and AI inference task flows. Pin 5: Custom serial communication TX pin, used to output sensor data and control commands; Pin 6: Custom serial communication RX pin, used to receive computing power scheduling instructions and inference results from the mobile terminal.

3. The cross-hardware AI memory interaction system according to claim 1, characterized in that, The hardware is adapted to a three-tier computing power collaborative architecture: Portable lightweight computing layer: It runs entirely on the built-in neural network processing unit of the core module and can run basic interactive functions offline; Mobile auxiliary computing power layer: The software agent calls on the mobile computing power, and the magnetic hardware only stores the agent's memory to execute medium-complexity inference tasks; Home-based heavy computing layer: When connected to a home global computing host, complex reasoning, long memory association, and multi-device scheduling tasks are offloaded to the host for execution, and the core module is only responsible for memory storage and interactive input and output; The three-tiered computing power collaborative architecture also includes a cloud-enhanced subsystem and a knowledge localization and accumulation mechanism: When connected to the internet, the API of a third-party cloud-based large model is called to process the de-identified common questions. The returned common knowledge fragments are extracted, vectorized, and written into a local encrypted storage unit, realizing knowledge accumulation that can be used permanently offline with a single call. The original privacy memory data is never fully uploaded to the public cloud.

4. The cross-hardware AI memory interaction system according to claim 1, characterized in that, The memory system in the encrypted storage unit adopts a three-layer architecture and supports cross-device encrypted synchronization: Working memory layer: Stores the current session context and is automatically cleared when the session ends; Contextual memory layer: Stores historical dialogues, event records, and emotion tags in the form of a vector database, supporting semantic retrieval; Semantic memory layer: Stores user preferences, behavioral habits, and thinking patterns in the form of structured data; The memory data is permanently stored in the core module and can be synchronized to the mobile parasitic agent and the home global computing power host. The synchronization process is encrypted throughout.

5. The cross-hardware AI memory interaction system according to claim 4, characterized in that, The memory system also includes: Time decay mechanism: Memory weights are calculated using an exponential decay function, and low-weight memories are automatically archived and compressed; Memory fusion mechanism: Periodically merge fragmented memories on the same topic into summary memories; Continuous learning mechanism: continuously update user features in the semantic memory layer based on user historical interactions.

6. The cross-hardware AI memory interaction system according to claim 1, characterized in that, It also includes building a dual-track agent ecosystem, including a digital software agent subsystem and a physical NFC shell agent subsystem; The digital software Agent subsystem includes an online Agent application store and a local Agent image library. Users can access the online store through the Agent software on their mobile terminals to browse, purchase, and download digital Agents with different functions. The downloaded Agent images are stored in the encrypted storage unit of the core module and support one-click switching of the software interface. The physical NFC shell Agent subsystem includes a replaceable NFC Agent shell, with a passive NFC tag embedded inside the shell, storing a unique authorized identifier for the Agent. After the core module reads the tag, it automatically loads the complete functional image of the corresponding Agent.

7. The cross-hardware AI memory interaction system according to claim 1, characterized in that, The Agent configuration set stored in the NFC tag includes: Agent Function Module ID: Points to the corresponding function plugin in the core module or home host, such as office assistant, life manager, sports coach, learning tutor, etc. Personality parameter set: Defines the tone of voice, word preferences, response rhythm, and logical tendencies of the agent's interaction; Voice packet identifier: points to the corresponding TTS speech synthesis model in the core module's storage unit; Permission configuration: Defines the memory levels and functional permissions that this Agent can access; The core module completes a full Agent switch within 1 second after reading the NFC tag. After the switch, the interaction logic, voice style, and available functions take effect simultaneously. Agents with different functions can be developed through the open SDK and distributed by binding customized shells.

8. The cross-hardware AI memory interaction system according to claim 1, characterized in that, It also includes a home host docking station, which is a fixed terminal for home use, including: Speaker arrays and circular microphone arrays are used for far-field voice interaction; High-speed synchronization interface: When the core module is placed on the base, an encrypted channel is automatically established to fully synchronize the memory data and the full Agent image library with the home global computing power host; Computing offload channel: forwards complex inference and multi-device scheduling tasks to home global computing power hosts for execution; Charging unit and independent auxiliary computing unit; Independent auxiliary computing unit: When the core module is not present, it can call upon the computing power of a home host to provide basic AI interaction services.

9. A cross-hardware AI memory interaction system according to claim 1, characterized in that, The core module also includes an NFC read / write module, which is used to read the NFC tag embedded in the replaceable shell and load the Agent image with the corresponding function.

10. A cross-hardware AI memory interaction system according to claim 1, characterized in that, The core module also includes an audio interaction unit, which includes a bone conduction vibration transducer and a drive circuit for achieving private voice output through contact conduction; and / or a microphone array for collecting user voice commands. The core module also includes a power management unit with a built-in lithium battery that supports wireless charging and PogoPin bidirectional power supply and dynamic power consumption adjustment. The system can be deployed in industrial production and enterprise office scenarios. The core module serves as the decision memory carrier for industrial robots or enterprise agents, storing production line operation logs, environmental decision data, and business process memories, and supports multiple terminals sharing a unified memory base.

Images