Large model memory management and interaction method based on dynamic cognitive map

Abstract

The invention discloses a large model memory management and interaction method based on a dynamic cognitive map. The interaction method comprises the following steps: receiving information input by a user; performing dialogue generation, intention recognition and information extraction by adopting a large language model inference engine to obtain an information recognition result; and performing dynamic memory management on the information identification result. And through contradiction detection and an active verification mechanism of the reflection engine, continuous pollution of wrong memory is effectively avoided, so that the system memory becomes more and more reliable over time.

Description

Technical Field

[0001] This invention relates to the field of long-term memory management, and more particularly to a large-scale model memory management and interaction method based on dynamic cognitive graphs. Background Technology

[0002] Currently, large models for long-term memory management in multi-turn dialogues often employ vector databases, key-value stores, or rule-driven memory update mechanisms. For example, existing solutions retrieve historical fragments in dialogues using Retrieval Augmentation Generation (RAG) technology, or manage information through hierarchical memory structures (long-term / short-term).

[0003] Existing technologies suffer from the following problems: memory is static and cannot be updated over time or with increased interaction; Isolated memories, lacking relevance, lead to contextual breaks; lack of credibility management makes it easy for erroneous memories to be continuously reinforced; the system responds passively and cannot actively initiate interactions based on memories.

[0004] The closest existing technology, such as the "multi-level memory bank" scheme, generates responses by retrieving different memory banks, but it still does not solve the problems of dynamic evolution of memory and active reasoning. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a large-scale model memory management and interaction method based on dynamic cognitive graphs to overcome or at least partially solve the above problems.

[0006] According to one aspect of the present invention, a large-scale model memory management and interaction method based on dynamic cognitive graphs is provided, the interaction method comprising: Receive information input by the user; A large language model inference engine is used for dialogue generation, intent recognition, and information extraction to obtain information recognition results; The information recognition results are dynamically stored and managed.

[0007] Optionally, the information received from user input includes text and language details.

[0008] Optionally, the dynamic memory management of the information recognition results specifically includes: Memory extractor: Extracts key information from the current dialogue and model output; Memory cell storage; Reflection and Verification Engine: An asynchronous background service that continuously scans, analyzes, and optimizes the database memory; Proactive cognitive interface: Based on dynamic cognitive graph, it performs reasoning, planning, and proactively initiates dialogue or services; Dynamic cognitive graph: Organize memory cells in the form of a knowledge graph, where nodes represent entities, concepts or assertions, and edges represent relationships; The knowledge graph is dynamically updated and forms the basis for the system to implement associative reasoning.

[0009] Optionally, the key information includes: entities, relationships, user intent, and sentiment.

[0010] Optionally, the relationship specifically includes: belonging, causing, and preference.

[0011] Optionally, the memory cell storage specifically includes: It stores structured memory cells, supporting efficient query, insertion, update, and deletion operations; it adopts a hybrid storage method combining graph databases and vector databases to support relational queries and semantic retrieval.

[0012] This invention provides a large-scale model memory management and interaction method based on dynamic cognitive graphs. The interaction method includes: receiving user input information; using a large language model inference engine to generate dialogue, identify intent, and extract information to obtain information recognition results; and performing dynamic memory management on the information recognition results. Through the contradiction detection and proactive verification mechanism of the reflection engine, the continuous contamination of erroneous memories is effectively avoided, making the system memory increasingly reliable over time.

[0013] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating a large-scale model memory management and interaction method based on dynamic cognitive graphs, provided in an embodiment of the present invention; Figure 2 A flowchart for memory lifecycle management provided in an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0017] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0019] This invention is applicable to scenarios that require long-term memory and contextual understanding, such as intelligent dialogue systems, personalized recommendation engines, virtual assistants, and medical consultation platforms.

[0020] like Figure 1 As shown, it includes the following components: User interaction interface: Receives user input (text, voice) and outputs system response.

[0021] Large Language Model (LLM) Inference Engine: Responsible for dialogue generation, intent recognition, and information extraction.

[0022] Dynamic memory management module: Memory lifecycle management flowchart as follows Figure 2 As shown.

[0023] ① Memory extractor: Extracts key information (entities, relationships, user intent, sentiment, etc.) from the current dialogue and model output.

[0024] ② Memory Cell Storage: Stores structured memory cells, supporting efficient query, insertion, update, and deletion operations. A hybrid storage approach combining graph databases (such as Neo4j) and vector databases (such as Milvus) can be used to support relational queries and semantic retrieval.

[0025] ③ Reflection and Verification Engine (Core Invention): An asynchronous background service that continuously scans, analyzes, and optimizes the memory.

[0026] ④ Proactive cognitive interface (core invention point): Based on dynamic cognitive graph, it performs reasoning, planning, and proactively initiates dialogue or services.

[0027] Dynamic cognitive graph: Memory cells are organized in the form of a knowledge graph, where nodes represent entities, concepts, or assertions, and edges represent relationships (such as "belongs to," "leads to," "preference"). The graph is dynamically updated and forms the basis for the system to achieve associative reasoning.

[0028] Example 1: Dynamic updates and proactive service for dietary preferences This embodiment details how the system processes the entire lifecycle of a user's specific memory of dietary preferences, including retrieval, storage, contradiction detection, verification, and active application. See the attached flowchart. Figure 2 : Memory lifecycle management flowchart.

[0029] Step S201 (Memory Retrieval and Creation): Trigger condition: The user says in the current conversation, "I've been really enjoying iced Americano lately." Execution entity: Memory extractor (collaborative LLM inference engine).

[0030] Processing actions: LLM identified "Iced Americano" as a beverage entity and "like" as a positive emotional tendency.

[0031] The memory extractor generates an initial "memory cell" based on a predefined template: Core assertion: Users like to drink iced Americano.

[0032] Evidence Source: {Source: Dialogue ID-20241014T1030; Type: User Explicit Statement; Original Statement: "I've been really into iced Americano lately."} Confidence level: 0.85 (The initial confidence level is high because it is a direct statement from the user) Timestamp: 2024-10-14 10:30:00 Decay factor: 360 (days) / / Dietary preference set to moderate decay rate. Status: Active Result: A new memory cell was created and stored in the memory cell storage device. At the same time, a node [User] - [Likes] -> [Iced Americano] was created in the dynamic cognitive graph.

[0033] Step S202 (Contradiction Detection and Confidence Management - Reflection Engine Operation): Triggering condition: The reflection engine scans the memory bank at preset intervals (e.g., every 24 hours).

[0034] Execution subject: Reflection and verification engine.

[0035] Processing actions: The engine retrieved a memory cell generated 3 months ago, whose core assertion is: The user reduced caffeine intake due to stomach discomfort. (Status: Dormant, Confidence: 0.90).

[0036] Through graph correlation analysis, the engine discovered a potential contradiction between "reducing caffeine intake" and "liking to drink iced Americano" (which contains caffeine).

[0037] The engine executes a contradiction detection algorithm to calculate the contradiction score of the two assertions. Since the score exceeds the threshold, the engine lowers the confidence level of the new memory cell (created in S201) from 0.85 to 0.65 and marks its status as "Pending Verification".

[0038] Result: The system doubts the accuracy of the memory and awaits an opportunity to verify it.

[0039] Step S203 (Active Verification - Active Cognition Interface Operation): Triggering condition: One day later, the user initiates a conversation and asks, "Any energy drinks you'd recommend today?" Execution entity: Active cognitive interface (collaborative LLM inference engine).

[0040] Processing actions: The proactive cognitive interface received a user query, retrieved relevant memories, and found the memory of "liking iced Americano" with the status of "pending verification" and the contradictory memory of "reducing caffeine".

[0041] The interface decision triggers a verification dialogue. LLM generates natural language statements: "I noticed you've been enjoying iced Americanos lately, but you previously mentioned reducing your caffeine intake due to stomach issues. I'd like to confirm if you're currently okay with caffeinated beverages so I can recommend more suitable drinks for you." The user replied: "Thank you for your concern. My stomach is better now, and I can drink coffee normally again." Result: The user's response served as new strong evidence.

[0042] Step S204 (Memory Update and Graph Evolution): Execution entities: memory extractor, memory cell storage device.

[0043] Processing actions: The memory extractor extracts information from the user's positive responses.

[0044] The system performs the following updates: ① Update the memory cell "likes to drink iced Americano": confidence level increased to 0.95, status restored to "active", and evidence source updated, adding this verification record.

[0045] ② Update the "Reduce Caffeine Intake" memory cell: the status has been changed to "Abandoned", and the reason for abandonment has been added: "User confirms change in health status".

[0046] ③ In the dynamic cognitive graph, strengthen the relationship weight of [User] - [Like] -> [Iced Americano], and weaken the old relationship.

[0047] Result: The memory bank and cognitive map were revised and evolved in a timely manner, becoming more accurate.

[0048] Step S205 (Proactive Service Recommendation): Execution subject: Active cognitive interface.

[0049] Processing actions: Based on verified memories and the user's current query ("energetic drink"), the interface is effective in inferring the user's current coffee preference from the cognitive graph.

[0050] The interface proactively suggests: "Based on your preferences, we recommend some of the highest-rated boutique coffee shops nearby serving iced Americanos. Also, if you'd like a change, drinks containing guarana are quite refreshing; would you be interested in trying them?" Result: The system provided highly personalized and thoughtful suggestions, improving the user experience.

[0051] Example 2 Personalized habit learning and proactive services in smart home scenarios In smart home integration scenarios, users control devices and express preferences through voice or text interaction with the smart hub. This system, acting as the "brain" of this smart hub, is responsible for learning user habits and providing a comfortable, energy-efficient, and intelligent home experience accordingly.

[0052] Memory retrieval: In an evening conversation, a user said, "Xiao Zhi, turn the living room air conditioner to 23 degrees Celsius and set the fan speed to the lowest setting. I'm going to read for a while and then go to sleep." The memory extractor, in conjunction with the LLM, identified the following key information: Behavioral Intent: Adjusting Air Conditioner Parameters Specific parameters: Temperature = 23℃, Wind force = Minimum Context: Time = Evening, User Activity = Reading, Subsequent Activity = Sleeping The system generates a "device preference" memory cell: Key assertion: When users go to sleep after reading in the evening, they prefer the living room air conditioner temperature to be 23°C and the fan speed to be at its lowest.

[0053] Evidence tracing: {Source: Dialogue ID-20241014T2230; Type: Explicit user instruction} Confidence level: 0.90 Timestamp: 2024-10-14 22:30:00 Decay factor: 180 (days) / / Habits may change slowly over time. Status: Active The memory cell is stored and associated with the nodes [User], [Evening], [Reading], [Sleeping], and [Living Room Air Conditioner] in the dynamic cognitive graph.

[0054] Reflection Engine Work (Habit Generalization and Conflict Detection): The reflection engine observed over the next few days that users issued the same or similar commands at similar times (22:15-22:45) on different evenings (Monday, Wednesday, and Friday).

[0055] Engine execution habit generalization: It recognized that this was a recurring pattern.

[0056] It creates a higher-order "habit" memory cell: Key assertion: Users have a habit of adjusting their living room air conditioner to 23°C and the lowest fan speed around 22:30 on weekday evenings (Monday to Friday), usually after reading activities.

[0057] Evidence tracing: {Source: Compilation of multiple dialogue records; Type: Systematic analysis and inference} Confidence level: 0.80 (based on pattern frequency and consistency) Status: Active Engine performs conflict detection: One Saturday night at 11:00 PM, a user gave the command: "Xiao Zhi, it's too cold, turn off the air conditioner." The engine detected a conflict between the instruction and the learned "habit" memory cells (the time was close but the instruction was opposite).

[0058] The engine doesn't simply assume the habit has failed; instead, it analyzes contextual differences: time = Saturday (non-working day), activity = unknown (the user didn't specify after reading and memorizing), user feedback = too cold.

[0059] The engine concludes that the habit is likely limited to weekdays. It modifies the assertion of the "habit" memory cell to more precisely that the user is on weekday evenings (Monday to Friday)... and associates the unusual event on Saturday as a new contextual constraint with the habit.

[0060] Proactive cognitive interface operation (proactive service and comfort assurance): Timing 1 (Habit Prediction and Execution): At 10:28 p.m. on a Tuesday night, the proactive cognitive interface, based on "habit" memory cells, predicted that the user would soon need to adjust the air conditioner.

[0061] The interface proactively initiates a conversation: "Based on your habits, I will now adjust the living room air conditioner to 23℃, low fan speed mode, is that alright?" The user replied: "Okay." The system performed the operation and reinforced the confidence level of the habit memory (from 0.80 to 0.85).

[0062] Timing 2 (Exception Handling and Personalized Care): One Wednesday evening, the outdoor temperature suddenly dropped. The system detected through the Internet of Things interface that the outdoor temperature had fallen below 18°C.

[0063] Active cognitive interface for reasoning: ① Habit: Users usually turn on the air conditioner at 23℃ at this time.

[0064] ② New context: It is very cold outside, and continuous heating may consume a lot of energy and cause discomfort.

[0065] The interface proactively intervened: "We noticed that it's particularly cold outside tonight. Are you still going to turn on the heating to 23℃ as usual? Or we suggest you try 22℃ first, which may be more comfortable and energy-efficient." After a user adopts a suggestion, the system generates a new "environmental adaptability" memory: when the outdoor temperature is below 18°C, the user can accept setting the living room air conditioner to 22°C. This will be used to optimize future proactive suggestions.

[0066] Seamless learning: The system automatically abstracts habit patterns from repeated user commands, reducing the hassle of repeated configuration for users.

[0067] Accurate predictions: It can distinguish the applicable conditions of habits (such as weekdays / weekends) and provide more accurate prediction services.

[0068] Active energy saving and comfort optimization: By combining real-time environmental data, we provide better and more energy-efficient personalized options while meeting user habits, reflecting the value of intelligence.

[0069] Continuous evolution: Habits are dynamically adjusted and optimized as user behavior changes, and the system has a true "learning" ability.

[0070] Beneficial effects: Improved memory accuracy: Through the reflection engine's contradiction detection and proactive verification mechanism, the continuous contamination of erroneous memories is effectively avoided, making the system's memory increasingly reliable over time.

[0071] Contextual depth understanding: The dynamic cognitive map establishes rich connections between memories, enabling the system to understand the causal relationship between "stomach disease recovery" and "restoration of coffee preference", thereby making inferences that are more in line with the user's current state.

[0072] A revolution in interactive experience: transforming passive responses into proactive care and service, the system can anticipate user needs (such as recommending beverages) and provide a highly humanized interactive experience.

[0073] Enhanced interpretability: Evidence traceability and state changes of memory cells record the "life course" of each important memory, making the system's decision-making process more transparent and interpretable to developers and users.

[0074] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A large model memory management and interaction method based on dynamic cognitive map, characterized in that, The interaction method comprises: Receiving user input information; Using a large language model inference engine for dialogue generation, intent recognition and information extraction to obtain information recognition results; Dynamic memory management is performed on the information recognition results.

2. The large model memory management and interaction method based on dynamic cognitive graph according to claim 1, characterized in that, The received user input information includes text and language details.

3. The large model memory management and interaction method based on dynamic cognitive graph according to claim 1, characterized in that, The dynamic memory management of the information recognition results specifically includes: Memory extractor: extracts key information from the current dialogue and model output; Memory cell storage; Reflection and verification engine: an asynchronous background service that continuously scans, analyzes and optimizes the memory library; Active cognitive interface: based on dynamic cognitive map, reasoning, planning and actively initiating dialogue or service; Dynamic cognitive map: organize memory cells in the form of knowledge graph, nodes represent entities, concepts or assertions, and edges represent relationships; The knowledge graph is dynamically updated and is the basis for the system to realize associative reasoning.

4. The large model memory management and interaction method based on dynamic cognitive graph according to claim 3, characterized in that, The key information includes: entity, relationship, user intent, sentiment orientation.

5. The large model memory management and interaction method based on dynamic cognitive graph according to claim 3, characterized in that, The relationship specifically includes: belongs to, causes, preference.

6. The large model memory management and interaction method based on dynamic cognitive graph according to claim 3, characterized in that, The memory cell storage specifically includes: Store structured memory cells to support efficient query, insert, update and delete operations; use a hybrid storage method combining graph database and vector database to support associative query and semantic retrieval.

Images