AI memory storage and retrieval system based on feature tag cross-hits

Abstract

This invention relates to the fields of artificial intelligence and information retrieval technology, and discloses an AI memory storage and retrieval system based on feature tag cross-hitting. This system constructs a seven-dimensional standard feature tag system (type, time, person, topic, location, importance, and sentiment). During storage, it automatically extracts 5-7 standard tags as metadata from the information. During retrieval, it decomposes the user's natural language query into tags and then uses cross-hit counting and sorting to accurately locate the target information. This invention also includes an experience solidification chain trigger mechanism, which writes valuable experiences from conversations into a skill base through a fully automated chain of scoring, confirmation, storage, verification, solidification, and recording. Compared with existing RAG technology, this invention does not rely on the embedding capabilities of AI models; tags are persistently stored in plain text; there is zero migration cost when changing models; retrieval results are transparent and interpretable; and it is closer to the natural memory methods of the human brain.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information retrieval technology, specifically to an AI memory storage and retrieval system based on feature tag cross-matching. Background Technology

[0002] Current mainstream AI systems suffer from the following fundamental problems in terms of memory and document retrieval:

[0003] (1) Pure filename or path dependency

[0004] Traditional file systems rely on users to remember filenames, paths, or directory structures to locate files. When users only remember the content characteristics of a file (such as "the Hokkaido skiing guide we discussed last time") rather than the filename, the retrieval efficiency drops sharply.

[0005] (2) Inherent limitations of RAG (Retrieval Enhancement Generation)

[0006] Current cutting-edge RAG technology employs vector embedding and semantic similarity retrieval, but suffers from the following problems:

[0007] A. High model coupling: Vector embedding depends on the embedding capability of a specific AI model. Changing the model requires re-embedding all the data.

[0008] B. Cross-model incompatibility: The vector spaces of different models are inconsistent, making direct transfer impossible;

[0009] C. Semantic ambiguity: Vector similarity cannot accurately distinguish the subtle differences between synonyms and near-synonyms;

[0010] D. High computational cost: Each query requires large-scale vector cosine similarity calculation, which consumes a lot of resources;

[0011] E. High cost of incremental updates: New files need to be re-embedded, which is not conducive to dynamic expansion;

[0012] (3) Lack of a structured labeling system

[0013] The existing system lacks unified and standardized label dimension definitions and minimum unit splitting rules, resulting in inconsistent label granularity and an inability to achieve stable and reliable cross-matching.

[0014] (4) Lack of experience consolidation mechanism

[0015] Valuable experience information acquired by AI in conversations is often lost when the conversation ends, lacking an automatic link from "conversation experience to structured tags, and then to solidification into a skill base".

[0016] Based on this, this application proposes an AI memory storage and retrieval system based on feature label cross-hitting to solve the above problems. Summary of the Invention

[0017] (a) Technical problems to be solved

[0018] To address the shortcomings of existing technologies, this invention provides an AI memory storage and retrieval system based on feature tag cross-hitting. This system constructs a seven-dimensional standard feature tag system. During storage, it automatically extracts 5-7 standard tags from the information as metadata. During retrieval, it decomposes the user's natural language query into tags and then uses cross-hitting counts and sorting to accurately locate the target information. At the same time, it adopts an experience solidification chain trigger mechanism to write valuable experiences in the dialogue into the skill library through a fully automated link of scoring, confirmation, storage, verification, solidification, and recording. It does not rely on the embedding capabilities of AI models. Tags are persistently stored in plain text, resulting in zero migration cost when changing models. The retrieval results are transparent and interpretable, and are closer to the natural memory methods of the human brain.

[0019] (II) Technical Solution

[0020] To achieve the above objectives, the present invention provides the following technical solution: an AI memory storage and retrieval system based on feature tag cross-hitting, comprising: a tag extraction module for extracting 5-7 standard feature tags for each piece of information to be memorized; a tag storage module for storing feature tags as metadata associated with the original information; a tag index library for storing all extracted feature tag sets; a query parsing module for receiving natural language queries from users and breaking them down into feature tags; a cross-hit counting and sorting module for calculating the number of hit tags for each stored piece of information and sorting them in descending order; and a result output module for outputting the information entry with the most hits.

[0021] Preferably, the feature tags include at least one of seven dimensions: type, time, person, theme, location, importance, and emotion.

[0022] Preferably, both the tag extraction module and the query parsing module include a compound word minimum unit splitting engine and a part-of-speech rule engine to ensure the atomicity and consistency of tag extraction.

[0023] Preferably, the compound word minimum unit splitting engine intelligently splits natural language into minimum units according to the following rules:

[0024] 1) Compound words must be broken down into the smallest independent noun or verb units;

[0025] 2) Proper nouns and fixed names retain their original forms.

[0026] Preferably, the part-of-speech rule engine rules are as follows:

[0027] Noun tags account for 60-70%, verb tags account for 20-30%, and adjective tags account for 0-10%.

[0028] The nouns include at least entity nouns of people, objects, or places; the verbs include at least actions, behaviors, or changes; and the adjectives are used only when expressing emotional dimensions or attributes.

[0029] Preferably, the query parsing module supports combined searches of any two or more tags. If any tag is matched, the information is considered to be related to the query. The more tag combinations, the higher the accuracy of the match. The sorting result of the cross-match counting and sorting module is white-box transparent and interpretable. The specific list of matched tags is displayed to the user each time a search is performed.

[0030] Preferably, it also includes an experience consolidation and chaining module, which comprises an experience scoring submodule, a repeat confirmation submodule, a tag storage submodule, a verification submodule, a skill consolidation submodule, and a process recording submodule, used to realize a fully automated chain from experience acquisition to skill consolidation; wherein:

[0031] The experience scoring submodule is used to score the quality of acquired experience and knowledge.

[0032] The repeat confirmation submodule is used to check whether similar knowledge already exists in the experience knowledge base;

[0033] The tag storage submodule is used to convert experiential knowledge into structured tags and add them to the tag index library;

[0034] The verification submodule is used to check the completeness and omissions of the experience-based solidification process;

[0035] The skills solidification submodule is used to write experience into a method skills library to form reusable methods and tools;

[0036] The process recording submodule is used to record the completion status of the entire fixed link;

[0037] Once the scoring result reaches a preset threshold, the experience scoring submodule automatically triggers all subsequent modules without manual intervention.

[0038] Another technical problem to be solved by this invention is to provide an AI memory storage and retrieval method based on feature label cross-hitting, comprising the following steps:

[0039] S1. Establish a multi-dimensional feature label system, which includes at least seven dimensions: type, time, person, theme, location, importance, and emotion.

[0040] S2. Extract 5-7 standard feature labels for each piece of information that needs to be stored. The feature labels are allocated according to the dimensions of the multi-dimensional feature label system.

[0041] S3. Store the extracted feature labels as metadata, associated with the original information.

[0042] S4. Receive the user's natural language query request and decompose the query request into multiple feature tags;

[0043] S5. Traverse the feature label set of all stored information in the system and calculate the number of hit labels for each stored information.

[0044] S6. Sort by the number of hit tags from highest to lowest, and return one or more information entries with the highest number of hits.

[0045] The above method does not rely on the embedding capabilities of AI models. Tags are persistently stored in plain text form. When changing AI models, there is no need to reprocess existing data. Tag retrieval and storage are based entirely on plain text matching. Vector embedding and cosine similarity calculation are not used, and the storage architecture does not require a vector database.

[0046] (III) Beneficial Effects

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This invention uses plain text tags for both storage and retrieval, eliminating the need for model embedding deployment and vector libraries (such as Pinecone / Chroma), thus reducing storage and computation costs. The tags themselves are persistent metadata, independent of any AI model, and the tags are retained without loss when changing models, enabling zero-cost migration across models.

[0049] 2. This invention uses tag set intersection counting sorting to achieve exact matching, which eliminates semantic ambiguity and is unaffected by semantic ambiguity. The hit accuracy is better than vector retrieval, and the retrieval precision is significantly improved. Moreover, the set intersection operation is faster than the cosine similarity calculation, especially in large data scenarios, where the retrieval speed is optimized. At the same time, the sorting results are transparent and interpretable. Each retrieval clearly shows which tags were hit, which is easy for users to understand. The more tags there are, the more accurate the retrieval. New tags can take effect immediately.

[0050] 3. This invention defines a seven-dimensional standard system of type, time, person, theme, location, importance, and emotion, covering all key dimensions of information, ensuring the consistency and completeness of tag extraction. At the same time, it defines a strict minimum unit splitting principle, that is, compound words must be split into independent minimum noun or verb units to ensure tag atomization, maximize cross-hit rate, and allow users to locate files based solely on natural language feature descriptions without having to remember file names, which can effectively improve user experience.

[0051] 4. This invention achieves automatic experience solidification through an automated experience solidification process, including scoring, confirmation, storage, verification, solidification, and recording. It is fully automatically triggered without human intervention, enabling automatic experience solidification. AI continues to learn and grow, and the entire solution completely simulates the natural memory and recall methods of humans. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0053] Figure 2 This is a schematic diagram of the storage and retrieval process of the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The difference between human memory mechanisms and AI lies in the fact that humans don't remember filenames, but rather features. For example, a person would say, "Last winter, my wife and I went to that hot spring hotel in Hokkaido," not "travel_2025_hokkaido_onsen.jpg." Current AI systems have completely deviated from this natural human way of remembering and retrieving information. Please see [link / reference]. Figure 1-2 This invention proposes an AI memory storage and retrieval system based on feature tag cross-hitting, which achieves a human-like file location and experience solidification mechanism through tag cross-hitting sorting. Specifically, it includes a tag extraction module, a tag storage module, a tag index library, a query parsing module, a cross-hit counting and sorting module, and a result output module. The tag extraction module extracts 5-7 standard feature tags for each piece of information to be memorized; the tag storage module stores the feature tags as metadata associated with the original information; the tag index library stores the set of all extracted feature tags; the query parsing module receives the user's natural language query and breaks it down into feature tags; the cross-hit counting and sorting module calculates the number of hit tags for each stored piece of information and sorts them in descending order; and the result output module outputs the information item with the highest number of hits.

[0056] The entire system uses tags for both access and retrieval, with feature descriptions driving the retrieval process. During operation:

[0057] Storage: User-described file features, AI automatically extracts multi-dimensional tags, and then the tags are used as metadata to bind to the file;

[0058] The process involves the user describing their needs in natural language, which the AI ​​then breaks down into feature labels. Finally, it iterates through all file metadata, performs cross-hit counts and sorts the results, and returns the matching files.

[0059] In this process, for each piece of information that needs to be memorized, the AI ​​automatically extracts 5-7 standard feature labels. The label dimensions are shown in the table below:

[0060] Dimension illustrate Example Type Information Categories Documents, conversation logs, images, links Time Time description January 2025, winter Person Relevant personnel Xiaoming, wife, colleague Theme core topics Hokkaido trip, code debugging, shopping list Location Geographical location Hokkaido, Niseko, Tokyo Importance Importance Important, Ordinary, Memo Emotion Emotional color Happy, regretful, recommended

[0061] The feature tag extraction and retrieval process employs the principle of minimum unit decomposition. Both the tag extraction module and the query parsing module include a compound word minimum unit decomposition engine and a part-of-speech rule engine to ensure the atomicity and consistency of tag extraction. The compound word minimum unit decomposition engine intelligently breaks down natural language into its smallest units, according to the following rules:

[0062] 1) Compound words must be broken down into the smallest independent noun or verb units;

[0063] 2) Proper nouns and fixed names retain their original forms.

[0064] For example: Incorrect example: "Niseko, Hokkaido", should be retained as a tag;

[0065] Correct examples: "Niseko, Hokkaido" should be split into "Hokkaido" and "Niseko" (two independent geographical tags); "Beijing Roast Duck Restaurant" should be split into "Beijing" and "Roast Duck" (noun splitting); "The Kite Runner" should retain the original title or proper noun.

[0066] Example of a long sentence: Input: "Last winter, my wife and I went skiing in Niseko, Hokkaido for three days"; Output: "last year", "winter", "wife", "Hokkaido", "Niseko", "skiing".

[0067] The rules for parts of speech are as follows:

[0068] Predominantly nouns: entity nouns such as people, objects, and places (accounting for 60-70%);

[0069] Verbs as secondary: action, behavior, change (accounting for 20-30%);

[0070] Use adjectives sparingly: use them only when expressing emotional dimensions or key attributes (0-10%).

[0071] The query parsing module supports searching with any combination of two or more tags. If any tag matches, the information is considered relevant to the query. The more tag combinations, the higher the accuracy. During query retrieval, the user enters a query statement, which the AI ​​breaks down into 5-7 independent tags and performs tag cross-matching on all stored files or entries in the system. An example is shown below:

[0072] 1) User query: "Help me find the document where I discussed skiing in Hokkaido with Xiaoming last time";

[0073] 2) AI-generated tag breakdown: "Xiaoming", "Hokkaido", "Skiing", "Document", "Last", "Discussion";

[0074] 3) Traverse all item tag sets and calculate the number of matching tags for each item;

[0075] 4) Sort by hit count in descending order;

[0076] 5) Return the entry with the most hits.

[0077] Single tag hit: All files containing "Xiaoming";

[0078] The two tags combined: "Xiaoming" ∩ "Hokkaido" will contain content related to Xiaoming in Hokkaido.

[0079] Multiple tag combinations: "Xiaoming" ∩ "Hokkaido" ∩ "Skiing" ∩ "Last time", accurately hit the target.

[0080] The more tags are stacked, the more accurate the results. It should be noted that the sorting results of the cross-hit counting and sorting module are white-box transparent and interpretable, and the specific list of hit tags is displayed to the user each time a search is performed.

[0081] This system also includes an experience consolidation and linkage module, which comprises an experience scoring submodule, a repeat confirmation submodule, a tag storage submodule, a verification submodule, a skill consolidation submodule, and a process recording submodule, used to realize a fully automated link from experience acquisition to skill consolidation; wherein:

[0082] The experience scoring submodule is used to score the quality of acquired experience and knowledge. If the score is ≥80, it will automatically continue.

[0083] The repeat confirmation submodule is used to check whether similar knowledge already exists in the experience knowledge base;

[0084] The tag storage submodule is used to convert experiential knowledge into structured tags and add them to the tag index library;

[0085] The verification submodule is used to check the completeness and omissions of the experience-based solidification process;

[0086] The Skills Solidification submodule is used to write experience into the method skill library to form reusable methods and tools (SKILL.md or tool scripts).

[0087] The process recording submodule is used to record the completion status of the entire fixed link and write it to the log;

[0088] Once the score reaches a preset threshold, the experience scoring submodule automatically triggers all subsequent modules without manual intervention.

[0089] This invention employs a plain text tag cross-matching retrieval method, which does not rely on any embedding model. The tag data is permanently reusable, and the model migration cost is zero. Computationally, it only involves set intersection operations, with a complexity far lower than vector retrieval. Incremental updates only require adding tag text, making maintenance convenient. This invention differs fundamentally from existing methods in terms of retrieval principles, model dependence, computational efficiency, and biomimicry, offering significant advantages such as precision and controllability, model independence, and high efficiency at low cost.

[0090] On the other hand, the present invention also proposes an AI memory storage and retrieval method based on feature label cross-hitting, comprising the following steps:

[0091] S1. Establish a multi-dimensional feature label system, which should include at least seven dimensions: type, time, person, theme, location, importance, and emotion.

[0092] S2. Extract 5-7 standard feature labels for each piece of information that needs to be stored. The feature labels are allocated according to the dimensions of the multi-dimensional feature label system.

[0093] S3. Store the extracted feature labels as metadata, associated with the original information.

[0094] S4. Receive the user's natural language query request and break it down into multiple feature tags;

[0095] S5. Traverse the feature label set of all stored information in the system and calculate the number of hit labels for each stored information.

[0096] S6. Sort by the number of hit tags from highest to lowest, and return one or more information entries with the highest number of hits.

[0097] The above method does not rely on the embedding capabilities of AI models. Tags are persistently stored in plain text form. When changing AI models, there is no need to reprocess existing data. Tag retrieval and storage are based entirely on plain text matching. Vector embedding and cosine similarity calculation are not used, and the storage architecture does not require a vector database.

[0098] The following examples illustrate the complete implementation of the present invention in different scenarios.

[0099] Example 1: AI Personal Assistant Memory System

[0100] User Faith has accumulated a large number of personal documents in her daily work and life, including travel guides, meeting minutes, shopping notes, code snippets, etc. Due to the large number and variety of files, the user often cannot accurately remember the name of each file and can only search based on a vague impression of the file's content characteristics. This embodiment demonstrates how this solution helps users quickly locate target files based on content descriptions.

[0101] I. File Storage

[0102] User Faith saved a PDF file titled "Hokkaido Winter Skiing Guide" into the system. After receiving the file, the system automatically analyzed the file content and contextual information, extracting and generating the following multi-dimensional tag system:

[0103] Type tag: #document;

[0104] Related tags: #wife;

[0105] Time tags: #January2025, #Winter;

[0106] Hashtags: #HokkaidoTravel, #Skiing;

[0107] Location tags: #Hokkaido, #Niseko;

[0108] Importance tag: #important;

[0109] Emotional tag: #expectation.

[0110] The aforementioned tags are bound to the PDF file as metadata and stored together in the tag index library. During the tag generation process, the system performs splitting processing on compound phrases. For example, "Niseko, Hokkaido" is split into two independent location tags: "Hokkaido" and "Niseko". "Skiing guide" is split into two independent topic tags: "Skiing" and "Guide". This splitting rule makes the tags more granular and allows for more flexible matching of various expressions by users during subsequent searches.

[0111] II. Document Retrieval

[0112] Some time later, user Faith wanted to retrieve the file but couldn't remember the filename. He entered a natural language query into the system: "Help me find the document about my wife and I going skiing in Hokkaido last winter." Upon receiving the query request, the system's query parsing engine first performed semantic analysis and tagging on the statement, extracting the following query tags: {#last year, #winter, #wife, #Hokkaido, #skiing, #document}. Then, the system traversed all file entries in the tag index, calculating the number of times each entry matched the above query tags, and sorting them from highest to lowest match count. The calculation showed that the file "Hokkaido Winter Skiing Guide.pdf" matched five tags: "winter," "wife," "Hokkaido," "skiing," and "document," ranking first in match count among all files, and therefore returned it as the first result to the user. Based on the initial search results, the user added a further limiting condition: "The document contains Niseko". The system then added the tag "Niseko" to the query tag set. At this point, the number of matching tags for the target file increased to 6, achieving a complete match of all query tags and accurately locking the search results to the relevant paragraphs in the file.

[0113] In this embodiment, the user does not need to recall or input a precise filename such as "2025_hokkaido_ski_guide.pdf". The file is located simply by describing the content characteristics of the file, such as "I went skiing in Hokkaido with my wife last winter". The cross-hatching and counting sorting mechanism of tags can transform the vague natural language description into precise matching logic, ensuring the accuracy and efficiency of the retrieval. At the same time, all tags are stored as independent metadata bound to the file. When the user changes the underlying artificial intelligence model, the established tag index system does not need to be regenerated, realizing the lossless preservation and cross-model reuse of tag assets.

[0114] Example 2: AI Dialogue Experience Consolidation Scenario

[0115] During long-term interactions with AI, the AI ​​often acquires valuable methodologies or operational experience when performing specific tasks. However, this experience usually only exists within the context of the current session and is cleared after the session ends. Even if the same or similar task requirements are encountered later, the AI ​​still needs to learn from scratch or rely on the user to repeatedly provide the same guidance information. This not only reduces the efficiency of interaction but also wastes experience resources. This embodiment adopts a solution that enables the valuable experience discovered by the AI ​​during task execution to be automatically extracted, structured and stored, and directly reused in subsequent sessions.

[0116] I. Experience Acquisition

[0117] A user requests the AI ​​to "help me write a script to automatically compile a weekly daily report." In processing this request, the AI, through interaction with the user and analysis of the task context, identifies and extracts the following reusable experiential information:

[0118] 1) The naming convention for daily report files is “YYYY_MM_DD_Daily Report_Name.md”;

[0119] 2) Daily report files are stored in the " / workspace / daily_reports / " directory;

[0120] 3) The core processing logic of the daily report is to extract the content of the "Key Outputs" and "Tomorrow's Plan" sections from the document.

[0121] The experience and information mentioned above are not only applicable to the current script writing task, but also provide guidance for other daily report processing needs that users may raise in the future.

[0122] II. Experience Consolidation Process

[0123] After recognizing the above experience, the AI ​​automatically triggers the experience solidification process, which is executed sequentially according to a preset ironclad chain mechanism:

[0124] 1) The system scores the practicality of the experience. After evaluation, the practicality score of this experience is 92 points, which reaches the preset automatic trigger threshold (≥80 points). Furthermore, after generalization analysis, it is determined that this experience can be reused in other similar scenarios involving the compilation of periodic reports that the user may raise in the future. Based on this, the system automatically enters the subsequent solidification process.

[0125] 2) The system checks the skill library for duplicates. After searching, it was found that there is no similar experience related to "Daily Report Document Processing" in the existing skill library. It was confirmed that the experience is new knowledge and needs to be stored as a new entry, rather than overwriting or merging existing entries.

[0126] 3) The system extracts experience content into a structured tagging system:

[0127] Type tag: #experience / workflow;

[0128] Related person tags: #No specific person;

[0129] Time-sensitive tag: #Still valid;

[0130] Tags: #Daily Report Compilation, #Document Processing, #Automation;

[0131] Path tag: # / workspace / daily_reports / ;

[0132] Importance tag: #important;

[0133] Emotional tag: #None;

[0134] The aforementioned tags are associated with the experience description documents as metadata and stored together in the tag index library;

[0135] 4) Based on this, the system performs integrity verification. After verification, the generated tag system fully covers all key elements such as daily report naming rules, storage path, and core processing logic, without omissions or deviations, and the verification is successful.

[0136] 5) In the final stage of experience consolidation, the system writes the specific operational logic of the experience into the skill file SKILL.md and generates a daily report organization function that can be directly called later, completing the transformation of experience from unstructured knowledge to an executable tool;

[0137] 6) After the entire curing process is completed, the system automatically records the completion status of the curing chain and the corresponding timestamp, and writes it into the operation log to ensure that the entire process is traceable.

[0138] Throughout the process, the AI ​​doesn't need to re-examine the naming rules, storage path, or processing logic of the daily report files. Instead, it automatically completes the following operations by retrieving pre-defined experience tags: locating the storage directory of the daily report files, matching the file naming rules, and extracting the content of the "Key Outputs" and "Tomorrow's Plan" sections according to the established processing logic, directly outputting the compiled results. The AI ​​doesn't perform any new learning or ask the user for repetitive background information; it directly reuses previously solidified experience, significantly shortening task response time and improving interaction efficiency. Furthermore, this experience, as a structured asset independent of specific conversations, is stored long-term in the skill library. Even if the user changes the AI ​​model used for interaction or starts a completely new conversation, as long as the skill library is loaded, this experience can be immediately reused without needing to be re-acquired or retrained.

[0139] Example 3: Cross-model transfer and multi-device collaboration scenarios

[0140] In practical use, users often don't rely on a single AI model or platform. For example, users might use multiple large language models simultaneously due to different task requirements, device limitations, or performance differences between service providers. However, different models typically maintain independent memory systems, meaning that user experience and stored personal information accumulated on model A cannot be directly accessed on model B. When switching models, users often need to provide the same background information again, leading to a fragmented user experience and reduced efficiency. This embodiment employs a solution to ensure seamless sharing of user memory assets across different models and devices.

[0141] I. Cross-model migration implementation process

[0142] First, during daily use of a general-purpose large language model (hereinafter referred to as "Model A"), users have gradually accumulated and stored hundreds of personal memory information entries. Each memory entry is extracted with five to seven tags according to the method in this scheme. All tags are stored in a structured JSON format as an independent plain text file, forming the user's personal tag index library. A partial example structure of this index library is as follows:

[0143] Memory entry 1: ID mem_001, corresponding file is "Hokkaido Winter Skiing Guide.pdf", tag set is {document, 2025, winter, wife, Hokkaido travel, Hokkaido, Niseko, skiing, important, looking forward to}, creation time is January 15, 2025, content summary is "Hokkaido Winter Skiing Guide";

[0144] Memory entry two: ID mem_002, corresponding file is "Meeting Minutes_0305.txt", tag set is {document, 2025, March, colleagues, project meeting, important, tense}, creation time is March 5, 2025, content summary is "Q1 project review meeting minutes";

[0145] The aforementioned tag library files exist independently of Model A and belong to the user's own digital assets, rather than being private data that cannot be transferred within Model A.

[0146] When a user decides to switch to a different platform's large language model (hereinafter referred to as "Model B"), the user only needs to copy the JSON file of the tag library into Model B's working environment. After Model B reads the file, it completely inherits the tag index of all hundreds of memories without any format conversion, model adaptation, or reprocessing. The user can immediately perform tag searches on Model B, such as searching for combinations of conditions like "Hokkaido + Skiing" or "Project Meeting + March". The tag cross-hitting ranking algorithm used in this solution is also effective on Model B, and the search results and ranking logic are completely consistent with Model A. Subsequently, any new memories added by the user on Model B are directly appended to the same tag library file after tag extraction, achieving continuous accumulation of memory assets and smooth model switching.

[0147] II. Implementation Process of Multi-Terminal Sharing

[0148] In addition to cross-model usage, users also need to access the same memory bank on multiple devices. This embodiment further demonstrates the specific implementation method of multi-terminal sharing, taking the user's usage scenario on different devices as an example:

[0149] Users deploy a master copy of the tag library on their personal servers. This master copy is automatically backed up daily and serves as the core storage and sole authoritative data source for the stored assets. Users' mobile devices periodically synchronize tag library files from the master library to their local devices for quick retrieval in daily mobile scenarios. Users' desktop devices pull the latest version of the tag library from the master library when starting work for desktop use.

[0150] Regarding the synchronization mechanism, each terminal runs the tag cross-hit sorting algorithm independently. Local retrieval and sorting do not depend on the server's real-time response, and can be used normally even when offline. When a user adds a memory tag on a terminal, the added data is periodically merged into the main database. If multiple places modify the same tag entry at the same time, the system adopts a conflict resolution strategy that prioritizes the latest timestamp to ensure that the data on each terminal is eventually consistent.

[0151] To more intuitively demonstrate the technical advantages of this embodiment, the following comparison will show the performance of this solution with commonly used vector embedding-based retrieval enhancement generation solutions in cross-model transfer scenarios.

[0152] In traditional vector retrieval schemes, if a user wants to migrate from one model to another, they need to go through the following steps: first, export all source documents, then recalculate vector embeddings for all documents on the new model, and since the embedding models of different models are different, the retrieval parameters need to be adjusted one by one. Finally, it is necessary to verify whether the retrieval quality has changed. The whole process may take several hours or even several days.

[0153] With this solution, users only need to export the tag library JSON file and copy it to the new model environment to complete the migration. No recalculation or parameter adjustment is required, and the retrieval quality is completely consistent with that before the migration. The whole process only takes a few seconds.

[0154] Therefore, this solution fundamentally solves the problems of repetitive computation and quality consistency in cross-model migration by decoupling the memory index information from the underlying model and treating the tag library as a standardized, lightweight asset independent of the model. This provides users with the technical foundation for free model selection and continuous memory accumulation. Furthermore, the JSON plain text format of the tag library is naturally suitable for multi-device synchronization, requiring no dedicated database or server support. Any device capable of reading text files can participate in the collaboration, significantly reducing the complexity and cost of multi-device deployment.

[0155] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI memory storage and retrieval system based on feature tag cross-hitting, characterized in that, include: The tag extraction module is used to extract 5-7 standard feature tags for each piece of information that needs to be memorized; The tag storage module is used to store feature tags as metadata associated with the original information. A tag index library is used to store the set of all extracted feature tags; The query parsing module is used to receive users' natural language queries and break them down into feature labels; The cross-hit counting and sorting module is used to calculate the number of hit tags for each stored piece of information and sort them in descending order; The results output module is used to output the information item with the highest number of hits.

2. The AI ​​memory storage and retrieval system based on feature tag cross-hitting according to claim 1, characterized in that, The feature tags include at least one of seven dimensions: type, time, person, theme, location, importance, and emotion.

3. The AI ​​memory storage and retrieval system based on feature tag cross-hitting according to claim 1, characterized in that, Both the tag extraction module and the query parsing module include a compound word minimum unit splitting engine and a part-of-speech rule engine to ensure the atomicity and consistency of tag extraction.

4. The AI ​​memory storage and retrieval system based on feature tag cross-hitting according to claim 3, characterized in that, The compound word minimum unit splitting engine intelligently splits natural language into the smallest units, according to the following rules: 1) Compound words must be broken down into the smallest independent noun or verb units; 2) Proper nouns and fixed names retain their original forms.

5. The AI ​​memory storage and retrieval system based on feature tag cross-hitting according to claim 3, characterized in that, The part-of-speech rule engine rules are as follows: Noun tags account for 60-70%, verb tags account for 20-30%, and adjective tags account for 0-10%. The nouns include at least entity nouns of people, objects, or places; the verbs include at least actions, behaviors, or changes; and the adjectives are used only when expressing emotional dimensions or attributes.

6. The AI ​​memory storage and retrieval system based on feature tag cross-hitting according to claim 1, characterized in that, The query parsing module supports searching by combination of any two or more tags. If any tag is matched, the information is considered to be relevant to the query. The more tag combinations, the higher the accuracy of the match. The sorting result of the cross-match counting and sorting module is white-box transparent and interpretable. The specific list of matched tags is displayed to the user each time a search is performed.

7. The AI ​​memory storage and retrieval system based on feature tag cross-hitting according to claim 1, characterized in that, It also includes an experience consolidation and chaining module, which comprises an experience scoring submodule, a repeat confirmation submodule, a tag storage submodule, a verification submodule, a skill consolidation submodule, and a process recording submodule, used to realize a fully automated chain from experience acquisition to skill consolidation; wherein: The experience scoring submodule is used to score the quality of acquired experience and knowledge. The repeat confirmation submodule is used to check whether similar knowledge already exists in the experience knowledge base; The tag storage submodule is used to convert experiential knowledge into structured tags and add them to the tag index library; The verification submodule is used to check the completeness and omissions of the experience-based solidification process; The skills solidification submodule is used to write experience into a method skills library to form reusable methods and tools; The process recording submodule is used to record the completion status of the entire fixed link; Once the scoring result reaches a preset threshold, the experience scoring submodule automatically triggers all subsequent modules without manual intervention.

8. An AI memory storage and retrieval method based on feature label cross-hitting, characterized in that, Includes the following steps: S1. Establish a multi-dimensional feature label system, which includes at least seven dimensions: type, time, person, theme, location, importance, and emotion. S2. Extract 5-7 standard feature labels for each piece of information that needs to be stored. The feature labels are allocated according to the dimensions of the multi-dimensional feature label system. S3. Store the extracted feature labels as metadata, associated with the original information. S4. Receive the user's natural language query request and decompose the query request into multiple feature tags; S5. Traverse the feature label set of all stored information in the system and calculate the number of hit labels for each stored information. S6. Sort by the number of hit tags from highest to lowest, and return one or more information entries with the highest number of hits.

9. The AI ​​memory storage and retrieval method based on feature label cross-hitting according to claim 8, characterized in that, The method does not rely on the embedding capabilities of the AI ​​model. The tags are persistently stored in plain text form, and there is no need to reprocess the existing data when changing the AI ​​model.

10. The AI ​​memory storage and retrieval method based on feature label cross-hitting according to claim 8, characterized in that, The method described above is based entirely on plain text matching for tag retrieval and storage, without using vector embedding or cosine similarity calculation, and the storage architecture does not require a vector database.

Images