CN-122019704-A - Intelligent response method, device, equipment and medium based on memory system drive
Abstract
The application discloses an intelligent response method, device, equipment and medium based on memory system driving. The method comprises the steps of responding to an input query statement, encoding the query statement into a context vector by combining current context information, searching a group of memory units related to the context vector in a dynamic memory knowledge graph in a multi-dimensional mode based on the context vector, constructing an association relation among the group of memory units based on a preset memory association analysis rule to obtain a memory subgraph, and carrying out multi-mode fusion on memory fragments corresponding to the group of memory units based on the memory subgraph to generate a natural language response. Through the mode, the memory can be characterized as the memory unit and stored in the dynamic memory knowledge graph, so that the memory capability of the intelligent body is greatly enhanced, the most relevant memory unit is quickly and accurately searched in the dynamic knowledge graph, and the accurate and consistent natural language response is generated through multi-modal fusion, so that the response efficiency and the response quality are improved.
Inventors
- JIAN WEIDONG
- GUO JIANLIN
- WANG CHAO
Assignees
- 深圳市有方科技股份有限公司
Dates
- Publication Date
- 20260512
- Application Date
- 20251231
Claims (11)
- 1. An intelligent response method based on memory system driving is characterized by comprising the following steps: Responsive to an input query statement, encoding the query statement in combination with current context information into a context vector; Based on the context vector, multi-dimensionally retrieving a set of memory cells associated with the context vector in a dynamic memory knowledge graph; Based on a preset memory association analysis rule, constructing a group of association relations among the memory units to obtain a memory subgraph; based on the memory subgraph, performing multi-modal fusion on the memory segments corresponding to a group of memory units to generate natural language response.
- 2. The intelligent response method based on memory system driving according to claim 1, wherein the multi-dimensionally retrieving a set of memory cells related to the context vector in a dynamic memory knowledge graph based on the context vector comprises: For the memory units in the dynamic memory knowledge graph, calculating a comprehensive relevance score based on one or more preset relevance evaluation factors; Based on the integrated relevance score, in combination with at least one search dimension independent of the integrated relevance score, a set of the memory cells associated with the context vector is searched from the dynamic memory knowledge-graph.
- 3. The intelligent response method based on memory system driving according to claim 1, wherein the memory unit is constructed as follows: And fusing the characteristics of each mode and atomizing the characteristics into the memory unit at least comprising a semantic tag and a time stamp.
- 4. The intelligent response method based on memory system driving according to claim 3, further comprising, after constructing the memory unit: The memory units are managed based on a hierarchical structure comprising multiple memory layers, importance of the memory units is calculated based on an importance evaluation algorithm, dynamic migration is carried out on the memory units among the multiple memory layers based on the importance, and the memory units which do not reach preset level standards corresponding to the corresponding memory layers are forgotten based on an adaptive forgetting mechanism and the importance.
- 5. The intelligent response method based on memory system driving according to claim 3, further comprising, after constructing the memory unit: Calculating a first association strength between the current memory unit and each memory unit in the dynamic memory knowledge graph based on the semantic tag and the timestamp; And when the first association strength meets a first preset association condition, establishing the association relation between the current memory unit and the corresponding memory unit in the dynamic memory knowledge graph.
- 6. The intelligent response method based on memory system driving according to claim 5, wherein the constructing a set of association relations between the memory units based on a preset memory association analysis rule to obtain a memory subgraph includes: calculating a second association strength between a set of the memory units based on the semantic tags, the time stamps, and the first association strength; and constructing the association relation between the corresponding memory units in the dynamic memory knowledge graph to obtain the memory subgraph when the second association strength meets a second preset association condition.
- 7. The intelligent response method based on memory system driving according to claim 1, wherein the generating a natural language response by multi-modal fusion of memory segments corresponding to a group of the memory units based on the memory subgraph includes: Based on a preset time line reconstruction method and a preset semantic coherent processing method, sequencing memory segments corresponding to each memory unit in the memory subgraph to obtain coherent memory segments; And carrying out multi-modal fusion on the coherent memory fragments, and generating the natural language response based on a pre-training language model.
- 8. The intelligent response method based on memory system driving according to claim 5, further comprising: Acquiring the retrieval frequency and the activation time corresponding to each memory unit, and performing memory conflict detection on the new original input information; Updating the importance of the memory unit in response to the search frequency of the memory unit being higher than a preset search frequency; updating the first association strength between a plurality of the memory cells in response to the activation times of the plurality of the memory cells being the same; In response to detecting the memory conflict, the dynamic memory knowledge graph is updated based on a preset memory correction mechanism.
- 9. An intelligent answering device based on memory system driving, comprising: The encoding module is used for responding to an input query statement and encoding the query statement into a context vector in combination with the current context information; a retrieval module for retrieving a set of memory cells associated with the context vector in a dynamic memory knowledge graph in multiple dimensions based on the context vector; The association module is used for constructing a group of association relations among the memory units based on a preset memory association analysis rule to obtain a memory subgraph; And the generation module is used for carrying out multi-mode fusion on the memory fragments corresponding to one group of memory units based on the memory subgraph to generate natural language response.
- 10. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to implement the steps of the memory system driven based intelligent answering method according to any one of claims 1-8.
- 11. A storage medium having stored thereon a computer program, which when executed by a processor performs the steps of the memory system driven based intelligent response method according to any of claims 1-8.
Description
Intelligent response method, device, equipment and medium based on memory system drive Technical Field The application relates to the technical field of artificial intelligence, in particular to an intelligent response method, device, equipment and medium based on memory system driving. Background With the rapid development of artificial intelligence technology, large language models (Large Language Model, LLM) have achieved remarkable results in the fields of natural language processing, knowledge questions and answers, etc., but existing AI agents generally lack persistent, structured memory capabilities. In the application scenarios of intelligent assistants, personalized services, educational coaching, medical diagnostics, etc., AI agents need to be able to remember historical interaction content, learn user preferences, and accumulate domain knowledge, and provide more personalized, consistent services based on these memories. The existing AI memory related technology mainly comprises: Context window based memory techniques-short term memory is maintained in a single session, but cannot be persisted across sessions, depending on the input length limitations of the model. External knowledge base enhancement technology, namely, historical information is stored through vector database and other technologies, but a structural organization and dynamic updating mechanism for memory contents is lacked. The prompting engineering method is characterized in that historical information is compressed into prompting words to be injected into a model, the prompting words are limited by the input length of the model, and dynamic evolution of memory cannot be realized. Disclosure of Invention The application mainly provides an intelligent response method, device, equipment and medium based on memory system driving, which are used for solving the problems that the memory capacity is insufficient, the memory retrieval efficiency and accuracy are low, and accurate response is difficult to generate in the existing AI memory technology. In order to solve the technical problems, the application adopts a technical scheme that an intelligent response method based on a memory system drive is provided. The method comprises the following steps: Responsive to an input query statement, encoding the query statement in combination with current context information into a context vector; Based on the context vector, multi-dimensionally retrieving a set of memory cells associated with the context vector in a dynamic memory knowledge graph; Based on a preset memory association analysis rule, constructing a group of association relations among the memory units to obtain a memory subgraph; based on the memory subgraph, performing multi-modal fusion on the memory segments corresponding to a group of memory units to generate natural language response. In an optional implementation manner of the embodiment of the present application, the multi-dimensional retrieval of a set of memory units related to the context vector in a dynamic memory knowledge graph based on the context vector includes: For the memory units in the dynamic memory knowledge graph, calculating a comprehensive relevance score based on one or more preset relevance evaluation factors; Based on the integrated relevance score, in combination with at least one search dimension independent of the integrated relevance score, a set of the memory cells associated with the context vector is searched from the dynamic memory knowledge-graph. In an optional implementation manner of the embodiment of the present application, the construction process of the memory unit includes: And fusing the characteristics of each mode and atomizing the characteristics into the memory unit at least comprising a semantic tag and a time stamp. In an alternative implementation manner of the embodiment of the present application, after the memory unit is constructed, the method further includes: The memory units are managed based on a hierarchical structure comprising multiple memory layers, importance of the memory units is calculated based on an importance evaluation algorithm, dynamic migration is carried out on the memory units among the multiple memory layers based on the importance, and the memory units which do not reach preset level standards corresponding to the corresponding memory layers are forgotten based on an adaptive forgetting mechanism and the importance. In an alternative implementation manner of the embodiment of the present application, after the memory unit is constructed, the method further includes: Calculating a first association strength between the current memory unit and each memory unit in the dynamic memory knowledge graph based on the semantic tag and the timestamp; And when the first association strength meets a first preset association condition, establishing the association relation between the current memory unit and the corresponding memory unit in the dynamic memory kno