A Multi-Agent Intellectual Property Processing System and Method Based on Temporal Memory Tree
By constructing a multi-agent intellectual property processing system with a five-layer temporal memory tree, the problems of memory fragmentation and fixed retrieval strategies in existing technologies are solved. This enables the automatic accumulation of user habits and industry experience, improves the accuracy and personalization of intellectual property business Q&A, and ensures the professionalism and efficiency of responses.
Patent Information
- Application Number
- CN202611435569.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing large-scale intellectual property models suffer from fragmented memory in scenarios such as patent agency, response to patent examination opinions, prior art retrieval, and case analysis. This results in the inability to achieve multi-level information management, leading to a single memory granularity, memory loss upon session termination, fixed search strategies, and the inability to automatically accumulate user habits and industry examination experience, resulting in insufficient professionalism and personalization in responses.
Design a multi-agent intellectual property processing system based on temporal memory trees. Construct a five-layer temporal memory tree, including L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer, and L5 global strategy layer. By automatically updating prompt words input into a large language model, hierarchical triggering and adaptive memory retrieval are achieved, improving the accuracy and personalization of business question answering.
It enables hierarchical storage and automatic updating of user interaction behavior, and adaptively selects the memory level according to the difficulty of the question, which improves the accuracy and personalization of intellectual property business Q&A, ensures that the answers are in line with user habits and case history, and improves the professionalism and efficiency of patent examination.
Smart Images

Figure CN123115398A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and intelligent intellectual property processing technology, and in particular to a multi-agent intellectual property processing system and method based on temporal memory trees. Background Technology
[0002] With the increasing prevalence of large language models in patent agency, patent examination opinion responses, prior art retrieval, and case analysis scenarios, existing intellectual property large model applications generally suffer from fragmented contextual memory. Traditional solutions only maintain memory for a single session window, which is lost once the session ends. This makes it impossible to accumulate data throughout the entire case lifecycle, user long-term work habits, or to summarize and form industry-wide examination strategies.
[0003] The existing technical solutions have the following main drawbacks: First, the memory granularity is too limited to distinguish between multi-level information such as interactive atomic behaviors, conversations, cases, personal habits, and industry-wide information; second, all questions use a unified retrieval scope, resulting in redundant retrieval for simple questions and incomplete memory recall for complex questions, leading to low reasoning efficiency; third, there is a lack of a hierarchical automatic update mechanism, relying on manual compilation of historical case experience, and failing to automatically accumulate user preferences and review domain patterns; fourth, it cannot achieve temporal memory management that aggregates at each level from "fine-grained interaction-conversation-case-user-global," and the output of large models is prone to being detached from the historical case context and not conforming to user work habits, resulting in insufficient professionalism and personalization in the responses.
[0004] Therefore, there is an urgent need to design a multi-agent intellectual property processing system and method based on temporal memory trees to solve the above problems. Summary of the Invention
[0005] This invention aims to overcome the shortcomings of existing large-scale intellectual property processing systems, such as single memory level, easy memory loss, fixed retrieval strategies, inability to automatically accumulate cases, and limitations in understanding user habits and industry examination experience. It provides a multi-agent intellectual property processing system and method based on temporal memory trees, which realizes the automatic construction of a five-layer temporal memory tree, layer-triggered updates, and adaptive selection of memory retrieval level according to the difficulty of the question, thereby improving the accuracy, personalization, and reasoning efficiency of intellectual property business Q&A.
[0006] The technical solution adopted by the embodiments of this application to solve its technical problem is: A multi-agent intellectual property processing system and method based on temporal memory trees, including a temporal memory tree construction module and a temporal memory tree retrieval module; The temporal memory tree construction module is used to independently maintain a five-layer temporal memory tree arranged from fine-grained to coarse-grained for each user. The five-layer temporal memory tree is, in order, L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer and L5 global strategy layer. The L1 atomic memory layer is used to generate time-ordered atomic memories in real time each time a user interacts with the system. The L2 session layer is used to integrate all L1 atomic memories within a single session to generate a session summary memory after the session ends. The L3 case layer is used to collect multiple L2 session memories of a single patent application to form a case memory. The L4 user habit layer is used to extract stable user behavior patterns and usage preferences based on multiple L3 case layer memories of the same user. The L5 global strategy layer is used to aggregate the L4 user habit layer memories of all users to generate a general knowledge graph and strategy base for patent examination. Each of the five-layer temporal memory tree layers extracts information from historical interaction data by inputting corresponding prompt words into the large language model, and is automatically updated according to its respective triggering conditions; The temporal memory tree retrieval module performs the following steps: The system receives user input questions and uses a large language model to classify the questions as simple, medium, or difficult. Simple questions are retrieved from layers L1 and L5, medium questions from layers L1, L3, and L5, and difficult questions from all five layers from L1 to L5. Perform vector similarity search at the selected level to obtain the Top-k memories most similar to the user's question; The top-k memories are deduplicated, sorted by time, and filtered for relevance. The filtered memories are then added to the context of the large language model to generate the final business output.
[0007] Furthermore, in the Top-k memories, k is a positive integer, and the value of k is 5.
[0008] Furthermore, the content recorded in the L1 atomic memory layer includes user-input query statements, user-set search conditions, patent documents clicked by the user, user's judgment results on the similarity of prior art documents, and user-confirmed or modified draft defense paragraphs.
[0009] Furthermore, the L2 session summary memory includes the case number of this session, the type of review opinion, the main technical issues discussed, and the user's key judgments and selections; the L2 session layer deduplicates and integrates the L1 atomic memory.
[0010] Furthermore, the L3 case layer memory includes basic information about the case application, key points of previous review opinions, content of previous responses and argumentation strategies, and the current legal status of the case; the L3 case layer is updated when a new review opinion is received or a response is completed.
[0011] Furthermore, the L4 user habit layer memory includes the user's professional expertise, preference for the depth and scope of comparative document retrieval, preference for defense strategies, and historical successful or failed defense patterns. When the L3 case layer memory reaches a preset quantity threshold, the L4 user habit layer update is initiated.
[0012] Furthermore, the L5 global strategy layer memory includes common types of examination opinions and corresponding strategies in different technical fields, examiner examination style patterns, and inventiveness judgment rules for combinations of prior art documents under different technical topics.
[0013] Furthermore, the update cycle for the L3 case layer is one day, the update cycle for the L4 user habit layer is one week, and the update cycle for the L5 global strategy layer is one month.
[0014] Furthermore, the processing method includes: The steps for constructing a temporal memory tree are as follows: Collect user interaction behavior, automatically construct and maintain the corresponding L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer, and L5 global strategy layer for that user; The temporal memory tree retrieval steps are as follows: receive user input questions, use a large language model to determine the difficulty of the questions, select the memory level to be retrieved based on the determination results, perform vector similarity search within the selected level to obtain the top-k memories, perform deduplication, time sorting and relevance filtering on the memories, add the filtered memories to the context of the large language model, and output the intellectual property business processing results.
[0015] Furthermore, the storage format of the memory includes at least one of JSON format and MD format.
[0016] In the above technical solution, the multi-agent intellectual property processing system and method based on temporal memory tree and the high-strength and tough bolt and its preparation method provided by the present invention have the following beneficial effects: 1. Construct a five-layer sequential memory tree from fine to coarse to achieve layered storage of atomic interactions, conversations, case files, user habits, and industry-wide strategies, solving the problem of memory loss when large model conversations end and the inability to accumulate historical case experience.
[0017] 2. The retrieval memory level is adaptively selected according to the difficulty of the question. Simple questions require fewer retrieval levels to ensure response speed, while complex questions require full five-level recall to ensure complete information, balancing reasoning efficiency and professionalism of the answer.
[0018] 3. Each level has a dual update mechanism that combines event triggering with timed cycles, automatically extracting memories from interaction logs without the need for manual organization and archiving, and automatically accumulating personal case-handling habits and a general patent examination strategy library for the entire platform.
[0019] 4. After being processed by vector recall, deduplication, temporal sorting, and relevance filtering, the memory is injected into the context of the large model. The output results are consistent with the case history, adaptable to the user's case handling habits, and in line with patent examination practice, significantly improving the quality of patent defense and case analysis output. Attached Figure Description
[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0021] Figure 1 This is an architecture diagram of the multi-agent intellectual property processing system based on temporal memory tree according to the present invention; Figure 2 This is a schematic diagram of temporal memory tree retrieval, which is the multi-agent intellectual property processing method based on temporal memory tree of the present invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. In addition, for the sake of convenience, the terms "upper," "lower," "left," and "right" are equivalent to the upper, lower, left, and right directions of the accompanying drawings themselves, and the terms "first," "second," etc., are used for descriptive purposes and have no other special meaning.
[0023] The technical solution in this application is to solve the above problems, and the overall approach is as follows: Example 1: This example presents a multi-agent intellectual property processing system based on temporal memory trees, such as... Figure 1-2 As shown, it includes a temporal memory tree construction module and a temporal memory tree retrieval module; The temporal memory tree construction module is used to maintain a five-layer temporal memory tree arranged from fine-grained to coarse-grained for each user. The five layers of the temporal memory tree are L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer and L5 global strategy layer. The L1 atomic memory layer is used to generate time-ordered atomic memories in real time each time a user interacts with the system. The L2 session layer is used to integrate all L1 atomic memories within a single session to generate a session summary memory after the session ends. The L3 case layer is used to collect multiple L2 conversation memories of a single patent application to form a case memory. The L4 user habit layer is used to extract stable user behavior patterns and usage preferences based on multiple L3 case layer memories of the same user. The L5 global strategy layer is used to aggregate the L4 user habit layer memories of all users and generate a general knowledge graph and strategy base for patent examination. Each of the five temporal memory tree levels extracts information from historical interaction data by inputting corresponding prompt words into the large language model and updates automatically according to its respective triggering conditions; The temporal memory tree retrieval module performs the following steps: The system receives user input questions and uses a large language model to classify the questions as simple, medium, or difficult. Simple questions are retrieved from layers L1 and L5, medium questions from layers L1, L3, and L5, and difficult questions from all five layers from L1 to L5. Perform vector similarity search at the selected level to obtain the Top-k memories most similar to the user's question; The top-k memories are deduplicated, sorted by time, and filtered for relevance. The filtered memories are then added to the context of the large language model to generate the final business output.
[0024] Specifically, the system independently constructs a five-layer temporal memory tree for each user. The five layers, from fine-grained to coarse-grained, are: L1 Atomic Memory Layer, L2 Conversation Layer, L3 Case Layer, L4 User Habit Layer, and L5 Global Strategy Layer. The five layers aggregate, abstract, and accumulate at each level, with temporal correlation, feature inheritance, and independent updates between layers.
[0025] The L1 atomic memory layer is responsible for capturing every user interaction in real time and generating the smallest granularity interaction memory unit with a timestamp. After a single conversation ends, the L2 conversation layer automatically aggregates all atomic memories of this conversation, removes duplicates, and extracts a conversation-level summary. For each patent case, the system collects all historical conversation memories of that case to form an independent L3 case complete file memory. Based on long-term multi-case user behavior data, the L4 user habit layer automatically extracts stable user search preferences, response styles, and judgment habits. Finally, the L5 global strategy layer gathers all user case data across the platform and precipitates a general knowledge graph and standard response strategy library for patent examination that spans multiple domains and examination scenarios.
[0026] Each level of memory extracts prompt words through a pre-set large language model, automatically extracts effective technical features, business information, behavioral features and strategy information from the original interaction logs in a structured manner, and automatically updates incrementally according to the exclusive trigger conditions of each level, without the need for manual organization and archiving.
[0027] During the business retrieval and reasoning phase, the system automatically categorizes user questions by difficulty, matching different levels of memory retrieval combinations to different difficulty levels: simple questions only retrieve the finest-grained real-time atomic memory and global general strategies to ensure response speed; medium-difficulty questions combine atomic interaction details, case history records, and global strategies to ensure the answer is relevant to the case history; difficult and complex defense questions, as well as creative in-depth analysis questions, fully retrieve all five layers of memory to achieve comprehensive contextual enhancement of real-time behavior, conversation details, case history, user habits, and industry global rules, greatly improving the relevance, personalization, and professional accuracy of the large model's answer.
[0028] During the retrieval process, the system performs vector similarity matching at the corresponding level, recalls the top-k highly similar historical memories, and after deduplication, temporal sorting, and relevance filtering, injects the high-quality historical memories into the context window of the large model. This constrains the model output to provide professional intellectual property processing results that conform to user habits, fit the case history, and are adapted to conventional examination strategies. On the hardware side, the processor runs the program code within the memory to implement all the above module logic.
[0029] In the top-k memories, k is a positive integer, and the value of k is 5.
[0030] Specifically, the number of similar memories, k, recalled by vector similarity retrieval can be any positive integer and can be dynamically adjusted according to server computing power, business concurrency, and question-answering accuracy requirements. In the preferred embodiment of this invention, k is set to 5. Recalling 5 optimal similar memories can cover similar historical cases, similar review questions, similar defense approaches, and users' historical operating habits, avoiding one-sided responses due to insufficient information, while also avoiding contextual redundancy, reasoning stagnation, and information chaos and conflict caused by excessive recalled data. It is the optimal balance point between accuracy and efficiency. In other embodiments, k can also be set to values such as 3, 4, 6, and 10, all of which are within the protection scope of this invention.
[0031] The L1 atomic memory layer records the following: user-inputted query statements, user-defined search criteria, patent documents viewed by the user, user's judgment on the similarity of prior art documents, and user-confirmed or modified draft responses.
[0032] Specifically, the L1 atomic memory layer is the smallest memory unit of the system, recording all original user interactions one by one. This includes: various patent search, inquiry, and defense consultation queries entered by the user in real time; search conditions manually set by the user, such as search keywords, classification numbers, date ranges, and claim field filtering; all prior art documents and prior art patent literature actively clicked, viewed, and downloaded by the user; the user's manual judgment results on the scope of disclosure, technical feature overlap, and level of inventiveness of the prior art documents; and the final draft content after the user confirms, deletes, rewrites, and corrects the defense paragraphs, modified statements, and argumentation logic generated by the model. This layer completely retains the user's first-hand operation trajectory, providing the most original and reliable data source for upper-level conversation aggregation, case review, and habit extraction.
[0033] The L2 session summary memory includes the case number, review opinion type, main technical issues discussed, and key user judgments and choices for this session; the L2 session layer deduplicates and integrates the L1 atomic memory.
[0034] Specifically, after a single human-computer interaction session ends, the system automatically triggers the generation of L2 session layer memories. The L2 session layer performs redundancy removal, behavior merging, and key point extraction on all L1 atomic memories within this session to generate a structured session summary memory. The core content stored in the session summary includes: the patent case number corresponding to this session, the type of examination opinion handled in this session (novelty issue, inventiveness issue, lack of necessary technical features, unclear, unsupported, out of scope, etc.), the core technical issues discussed in this session, the key choices made by the user during the session, technical trade-offs, judgment conclusions, and modification preferences.
[0035] The L2 layer implements semantic abstraction from "scattered operation logs" to "single work intent", significantly compressing invalid data and retaining valid business information.
[0036] The L3 case layer memory contains basic information about the case application, key points of previous review opinions, revisions and argumentation strategies for each response, and the current legal status of the case; the L3 case layer is updated when new review opinions are received or a response is completed.
[0037] Specifically, the L3 case layer uses a single patent application as the sole aggregation dimension and is the core business memory layer of this invention. The case layer summarizes all L2 session summaries of the case since its application, forming a complete case lifecycle archive.
[0038] The L3 case-level memory specifically includes: basic information such as the case application number, application date, technical field, core solution, and claim layout; the key points of contention in each examination opinion, the examiner's angle of questioning, and the core reasons for rejection; the textual modifications, claim adjustments, technical argumentation logic, and defense strategies during each response process; and the real-time legal status of the case, including whether it is pending a response, has been responded to, has been rejected, has been granted, or has been withdrawn.
[0039] The L3 case layer has an event-driven update mechanism: when the system recognizes that the case has received new review comments or the user has completed a new round of defense revisions and submissions, it will immediately and incrementally update the case memory to ensure that the case file is up-to-date.
[0040] The L4 user habit layer memory includes the user's professional expertise, preference for the depth and scope of comparative document retrieval, preference for defense strategies, and historical success or failure patterns in defense. When the L3 case layer memory reaches a preset threshold, the L4 user habit layer update is initiated.
[0041] Specifically, the L4 user habit layer focuses on the long-term stable behavioral characteristics of individual users, stripping away the differences in single, accidental operations, and accumulating users' fixed work styles.
[0042] Its memory content includes: the technical professional fields that users frequently process over a long period of time, the database preferences of users searching for comparison documents, the depth preference of searching keywords (broad or detailed), the usual angles of argumentation, the tendency of claim amendment strategies, the common means of avoiding defects, the fixed logical patterns of historical successful arguments, and the failure characteristic patterns of arguments that fail and responses that are rejected.
[0043] The system is equipped with a quantity threshold trigger mechanism. When the accumulated L3 case memories collected by the user reach the preset threshold, batch data analysis and habit reconstruction and update will be automatically started to ensure that the user profile is objective, stable and effective.
[0044] The L5 global strategy layer memory includes common types of examination opinions and corresponding strategies in different technical fields, examiner examination style patterns, and the rules for judging the inventiveness of combinations of prior art documents under different technical themes.
[0045] Specifically, the L5 global strategy layer is a platform-level industry knowledge base that aggregates the accumulated data of all users and all historical cases to form a general patent examination and processing strategy library.
[0046] Its content includes: high-frequency examination opinion types, standard questioning logic, and mature response strategies in different technical fields such as mechanics, electrical engineering, optoelectronics, materials, medical devices, and algorithm software; the routine examination styles, questioning focuses, and rejection preferences of different examiners; and the general judgment rules for the combination of prior art documents, the logic for identifying distinguishing features, and the validity or invalidity of inventiveness assessments under different technical themes.
[0047] This level upgrades from "personal experience" to "industry-standardized experience," significantly improving the accuracy of initial responses to new users and new cases.
[0048] The update cycle for the L3 case layer is one day, the update cycle for the L4 user habit layer is one week, and the update cycle for the L5 global strategy layer is one month.
[0049] Specifically, the five-layer memory system adopts a dual mechanism of event-triggered updates and timed batch refreshes to ensure data real-time performance and integrity.
[0050] Among them, the L3 case layer has frequent business changes and is set to be updated once a day to uniformly organize all session and case change data of the day; the L4 user habit layer is a long-term behavior accumulation with little short-term change and is set to be updated once a week to batch iterate user behavior profiles; the L5 global strategy layer is a macro-level industry rule with slow data iteration and is set to be updated once a month to complete the iterative upgrade of the entire network strategy library. Layered and staggered updates reduce server pressure and ensure that the memory accuracy of each layer matches its own business characteristics.
[0051] Example 2: This example presents a multi-agent intellectual property processing method based on temporal memory trees, such as... Figure 1-2 As shown, the processing methods include: The steps for constructing a temporal memory tree are as follows: Collect user interaction behavior, automatically construct and maintain the corresponding L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer, and L5 global strategy layer for that user; The temporal memory tree retrieval steps are as follows: receive user input questions, use a large language model to determine the difficulty of the questions, select the memory level to be retrieved based on the determination results, perform vector similarity search within the selected level to obtain the top-k memories, perform deduplication, time sorting and relevance filtering on the memories, add the filtered memories to the context of the large language model, and output the intellectual property business processing results.
[0052] Specifically, the processing method relies on the aforementioned system to achieve automated and intelligent processing, and is divided into two core steps: The steps for constructing a temporal memory tree are as follows: The system continuously collects all user interaction behavior data, parses and extracts it layer by layer, automatically maintains and continuously iterates and updates the user's exclusive L1 to L5 five-layer temporal memory tree, and realizes the full life cycle of data accumulation.
[0053] The temporal memory tree retrieval reasoning steps are as follows: After receiving real-time business inquiries from users, the system first uses a large language model to intelligently determine the difficulty level of the question and matches a specific retrieval level combination; it then performs a similarity search in the corresponding level vector library to obtain the top-k high-quality historical memories; the recalled content is uniformly deduplicated, sorted chronologically, and filtered for relevance; all the filtered high-quality historical memories are injected into the context of the large language model, constraining the model to combine user personal habits, case history, and industry-standard strategies to output accurate, personalized, and compliant intellectual property business processing results.
[0054] The storage format of memory includes at least one of JSON and MD formats.
[0055] Specifically, all hierarchical time-series memory data are stored in at least one of the following formats: JSON and Markdown.
[0056] Structured fields, parameter data, tag data, and behavioral statistics are stored in JSON format for easy machine parsing, vector input, and automatic comparison and updating. Long text summaries, defense strategy texts, experience summaries, and case review content are stored in Markdown format for easy human reading, editing, and archiving. The system can store data in a single format or in parallel with both formats to adapt to the storage and retrieval needs of different data types.
[0057] This invention discloses a multi-agent intellectual property processing system and method based on a temporal memory tree, primarily applied to intelligent intellectual property business scenarios such as patent case retrieval, examination opinion analysis, response strategy generation, user behavior accumulation, and global examination experience summarization. By constructing a five-layer hierarchical temporal memory tree, this invention achieves hierarchical accumulation, temporal updates, and intelligent retrieval adaptation from fine-grained interactive atomic behaviors to global industry strategies. This addresses the technical shortcomings of traditional large-scale intellectual property question-and-answer models, including contextual forgetting, lack of case memory, lack of user habit adaptation, lack of domain examination pattern accumulation, single-granularity question-and-answer matching, and low retrieval efficiency across uniform difficulty scenarios.
[0058] The contents not described in detail in this specification are existing technologies known to those skilled in the art.
[0059] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A multi-agent intellectual property processing system based on temporal memory trees, characterized in that, This includes a temporal memory tree construction module and a temporal memory tree retrieval module; The temporal memory tree construction module is used to independently maintain a five-layer temporal memory tree arranged from fine-grained to coarse-grained for each user. The five-layer temporal memory tree is, in order, L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer and L5 global strategy layer. The L1 atomic memory layer is used to generate time-ordered atomic memories in real time each time a user interacts with the system. The L2 session layer is used to integrate all L1 atomic memories within a single session to generate a session summary memory after the session ends. The L3 case layer is used to collect multiple L2 session memories of a single patent application to form a case memory. The L4 user habit layer is used to extract stable user behavior patterns and usage preferences based on multiple L3 case layer memories of the same user. The L5 global strategy layer is used to aggregate the L4 user habit layer memories of all users to generate a general knowledge graph and strategy base for patent examination. Each of the five-layer temporal memory tree layers extracts information from historical interaction data by inputting corresponding prompt words into the large language model, and is automatically updated according to its respective triggering conditions; The temporal memory tree retrieval module performs the following steps: The system receives user input questions and uses a large language model to classify the questions as simple, medium, or difficult. Simple questions are retrieved from layers L1 and L5, medium questions from layers L1, L3, and L5, and difficult questions from all five layers from L1 to L5. Perform vector similarity search at the selected level to obtain the Top-k memories most similar to the user's question; The top-k memories are deduplicated, sorted by time, and filtered for relevance. The filtered memories are then added to the context of the large language model to generate the final business output.
2. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, In the Top-k memories, k is a positive integer, and the value of k is 5.
3. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, The L1 atomic memory layer records the following: user-inputted query statements, user-set search conditions, patent documents viewed by the user, user's judgment results on the similarity of prior art documents, and user-confirmed or modified draft defense paragraphs.
4. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, The L2 session summary memory includes the case number, review opinion type, main technical issues discussed, and key user judgments and selections for this session; the L2 session layer deduplicates and integrates the L1 atomic memory.
5. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, The L3 case layer memory includes basic information about the case application, key points of previous review opinions, content of revisions and argumentation strategies for previous responses, and the current legal status of the case; the L3 case layer is updated when new review opinions are received or a response is completed.
6. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, The L4 user habit layer memory includes the user's professional expertise, preference for the depth and scope of comparative document retrieval, preference for defense strategies, and historical successful or failed defense patterns. When the L3 case layer memory reaches a preset quantity threshold, the L4 user habit layer update is initiated.
7. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, The L5 global strategy layer memory includes common types of examination opinions and corresponding strategies in different technical fields, examiner examination style patterns, and inventiveness judgment rules for combinations of prior art documents under different technical themes.
8. The multi-agent intellectual property processing system based on temporal memory tree according to claim 1, characterized in that, The update cycle for the L3 case layer is one day, the update cycle for the L4 user habit layer is one week, and the update cycle for the L5 global strategy layer is one month.
9. A method for processing intellectual property rights based on a time-series memory tree in a multi-agent intellectual property processing system based on any one of claims 1-8, characterized in that, The processing method includes: The steps for constructing a temporal memory tree are as follows: Collect user interaction behavior, automatically construct and maintain the corresponding L1 atomic memory layer, L2 conversation layer, L3 case layer, L4 user habit layer, and L5 global strategy layer for that user; The temporal memory tree retrieval steps are as follows: receive user input questions, use a large language model to determine the difficulty of the questions, select the memory level to be retrieved based on the determination results, perform vector similarity search within the selected level to obtain the top-k memories, perform deduplication, time sorting and relevance filtering on the memories, add the filtered memories to the context of the large language model, and output the intellectual property business processing results.
10. The multi-agent intellectual property processing method based on temporal memory tree according to claim 9, characterized in that, The storage format of the memory includes at least one of JSON format and MD format.
