A vector embedding-based association inference method and system for personal memory data

Abstract

The application discloses a kind of vector embedding-based association inference method for personal memory data, comprising: converting personal memory data organized in structured format containing predefined semantic anchor point field (at least including time, place, person, dialogue, event, emotion) into vector embedding representation;Calculate the semantic similarity between the vector embedding and the stored memory vector embedding;Identify associated memories based on semantic similarity and output.The method covers wearable devices, mobile terminals, vehicle-mounted systems and smart home computing terminals, supports time-aware embedding and anchor point field weighted matching, and fills the protection gap of existing GNN memory association patents in "non-graph structure path".

Description

Technical Field

[0002] This invention belongs to the field of memory data retrieval and vector computing technology, specifically relating to a method and system for performing associative reasoning on personal memory data organized with predefined semantic anchor fields through vector embedding for wearable devices, mobile terminals, in-vehicle systems, and smart home devices.

[0003] Terminology Definition

[0004] The terms used in this specification and claims are defined as follows:

[0005] (1) Personal memory data: including but not limited to structured data, unstructured text, raw audio and video data, feature identification data, fingerprint data and any combination thereof, as long as its content or identification is related to the experience, behavior and status of a specific individual.

[0006] (2) Processing and calling: including but not limited to any computational behavior that can acquire, present or utilize memory data content, such as retrieval, matching, addressing, mapping, generation, reasoning, context window association, etc.

[0007] (3) Vector embedding representation: refers to mapping personal memory data into dense vector representation in a high-dimensional vector space. This mapping can be performed by an independent embedding model or generated by the underlying network layer of the edge inference model (including large language model, multimodal model, small Transformer model, etc.). Background Technology

[0008] Existing memory crosslinking patents primarily rely on Graph Neural Networks (GNNs) for association reasoning, constructing crosslinking graphs using memory data as graph nodes and calculating the degree of association between entities. However, attackers can circumvent GNN-based protection by employing non-GNN schemes—such as vector embedding + similarity retrieval. Vector retrieval schemes convert memory data into vector embeddings, using semantic similarity between vectors to replace GNN crosslinking reasoning, achieving functional equivalence without using a graph structure. As of June 2026, there are no patents covering the complete reasoning path of "converting personal memory data into vector embeddings → performing association retrieval through semantic similarity" for personal memory association. This patent closes this gap, ensuring that attackers fall within the protection scope regardless of whether they use GNNs or vector retrieval to achieve memory association. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a vector embedding-based association reasoning method for personal memory data, which fills the protection gap of existing GNN memory association patents on "non-graph structure paths".

[0010] This invention provides a method for associative reasoning of personal memory data, comprising: converting at least one piece of personal memory data into a vector embedding representation, wherein the personal memory data is organized in a structured format including predefined semantic anchor fields (including at least time, location, person, dialogue, event, and emotion); calculating the semantic similarity between the vector embedding representation and a plurality of stored memory vector embeddings; and identifying and outputting memory data associated with the personal memory data based on the semantic similarity.

[0011] The invention further includes a push step: when the semantic similarity exceeds a preset threshold, the associated memory data is pushed to the user interface. The output method of the user interface includes one or more of visual display, audio playback, and haptic feedback.

[0012] The present invention also provides a personal memory data association reasoning system based on vector embedding, including a vectorization module, a similarity calculation module, and an association output module.

[0013] Regarding the specific implementation of vector embedding generation, this invention supports, but is not limited to, the following solutions:

[0014] (1) Pre-trained embedding model: Generate semantic vectors of memory data based on general text embedding models (such as text-embedding-3, BGE, E5).

[0015] (2) Time-aware embedding model: Add a time encoding component to the basic embedding model to encode the collection timestamp of the memory data into a continuous time vector, so that memory data with similar timestamps are close to each other in the embedding space.

[0016] (3) Anchor field weighted embedding: The encoding of each dimension is weighted according to the importance weight of the predefined semantic anchor field. For example, the weight of the dialogue field is higher than that of the sentiment field, so that the association matching focuses more on semantic content rather than sentiment state.

[0017] Compared with existing technologies, this invention has the following advantages: First, it fills the protection gap in GNN memory association patents regarding non-graph structured reasoning paths. Second, it supports local vectorization on the device side, with only vector embeddings uploaded to the collaborating device for cross-device association. Third, it is naturally compatible with the F1 (data model) data structure—personal memory data is organized in a structured format with predefined semantic anchor fields, and vector embeddings are extracted from the structured data through an embedding model. Attached Figure Description

[0018] Figure 1 This is a flowchart of the personal memory data association reasoning method based on vector embedding of the present invention.

[0019] Figure 2This is a schematic diagram illustrating the vector embedding generation and similarity calculation of the present invention.

[0020] Figure 3 This is a user interaction flowchart for the push notification of related results in this invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the present invention.

[0022] Example 1: End-side vectorization and association retrieval

[0023] The smart glasses collected a new personal memory ("Discussing a project with Zhang San at a coffee shop on the afternoon of June 17, 2026"). After storing the data in a structured format according to predefined semantic anchor fields, the edge embedding model concatenated the anchor fields into a text sequence and converted it into a 256-dimensional vector embedding. The system calculated the cosine similarity between this vector embedding and the vector embeddings of 5,000 locally stored memory data. In the similarity ranking, the memory related to "Zhang San" had the highest similarity (0.92), followed by the memory related to "coffee shop" (0.85), and then the memory related to "June 2026" (0.78). The system returned the top-5 associated memories to the user interface.

[0024] Example 2: Time-Aware Vector Embedding

[0025] The embedding model includes a time-aware component: the Unix timestamps of the memory data are encoded into periodic time vectors (containing sine and cosine codes of four frequency components: hour, week, month, and year), and concatenated after the basic semantic embedding. When a user queries "what did I do last week", the query text is vectorized and its similarity is calculated with all memory vectors. The time encoding naturally gives last week's memory a higher similarity score.

[0026] Example 3: Weighted Matching of Anchor Fields

[0027] The weights of the "Dialogue" anchor field are set to 0.3, "Event" anchor field to 0.25, "People" anchor field to 0.2, "Time" anchor field to 0.15, and "Location" anchor field to 0.1. When a user queries "What was the new project discussed that day?", the weighting of the dialogue and event fields allows semantic matching to more accurately pinpoint memories related to "discussions about the new project".

[0028] Example 4: Cross-device vector association

[0029] When the smartwatch is offline, it collects a segment of memory data and performs vector embedding calculations locally. Once connected to the network, the smartwatch uploads the vector embedding (not the original data) to the cloud for synchronization. When a new related memory is generated on the user's phone, the cloud detects a high similarity (0.85 > threshold 0.75) between the new and the uploaded vector embedding, and automatically pushes the association result to the phone's user interface. The original memory data remains locally on the device that generated the data and is never uploaded to the cloud, protecting user privacy.

[0030] Example 5: Vector Embedding Generated by the Low-Level Network of a Large Model

[0031] A lightweight large language model (LLM) is deployed on the smart glasses. A user's personal memory data ("Having lunch with my wife in Yantian District on June 18, 2026") is fed into the LLM for processing. The LLM automatically generates a high-dimensional semantic representation of this memory data in its intermediate hidden layers (such as the hidden states before the last layer) of its Transformer architecture—this representation is essentially a vector embedding. The system directly reads the output vector of the LLM's intermediate hidden layers as the vector embedding representation of this memory, without using a separate embedding model. Subsequently, the system calculates the similarity between this vector and the vectors of other stored memory data to locate the relevant memory. In this embodiment, the vector embedding is generated by the LLM's underlying network; there is no independent vector embedding module. However, the generated vector representation still achieves the function of semantic association calculation, which is equivalent to the fact within the scope of this patent.

Claims

1. A method for associative reasoning of personal memory data, characterized in that, include: At least one segment of personal memory data is converted into a vector embedding representation, the personal memory data being organized in a structured format that includes predefined semantic anchor fields; Calculate the semantic similarity between the vector embedding representation and multiple stored memory vector embeddings; Based on the semantic similarity, memory data associated with the personal memory data is identified and output.

2. The method as described in claim 1, characterized in that, The method is applicable to at least one of wearable devices, mobile terminals, in-vehicle systems, smart home devices, and any computing device containing a processor and memory.

3. The method as described in claim 1, characterized in that, The vector embedding representation is generated based on a pre-trained embedding model that includes a time-aware component, enabling memory data with similar timestamps to obtain more similar vector representations.

4. The method as described in claim 1, characterized in that, The semantic similarity calculation includes at least one of cosine similarity, Euclidean distance, and dot product.

5. The method as described in claim 1, characterized in that, Further includes: When the semantic similarity exceeds a preset threshold, the associated memory data is pushed to the user interface.

6. The method as described in claim 5, characterized in that, The user interface output methods include one or more of visual display, audio playback, and haptic feedback.

7. The method as described in claim 1, characterized in that, The personal memory data is stored in a structured format that includes predefined semantic anchor fields, which include at least a time field, a location field, a person field, a dialogue field, an event field, and an emotion field.

8. The method as described in claim 1, characterized in that, The vector embedding is calculated locally on the device that generates the personal memory data, and only the vector embedding is uploaded to the collaborating device for cross-device memory association.

9. The method as described in claim 1, characterized in that, The vector embedding representation is generated by the underlying network layer of the edge inference model, without the need for a separate vector embedding module. The edge inference model includes at least one of a large language model, a multimodal model, and a small Transformer model.

Images