A method, system, electronic device, and storage medium for constructing, evolving, and retrieval the life memory of multi-source enterprise data.

By generating source identity identifiers and content fingerprints, and separating analysis processing from memory generation processing, the problem of repeated ingestion of multi-source data and historical state overwriting for enterprises is solved, enabling traceable data retrieval and permission consistency, and improving the accuracy and security of data processing.

CN123115131APending Publication Date: 2026-11-17NANJING HUAXI YUNKE INTELLIGENT TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611139584.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Enterprises face problems such as duplicate data intake, difficulty in recovery from interruptions, overwriting of historical states, untraceable sources, unstable mixed retrieval, and unauthorized reuse of cached data after permission changes during long-term accumulation.

Method used

By generating source identity identifiers and content fingerprints, separating analysis and processing from memory generation, establishing memory units and using permission snapshots to filter candidate memory units, a chain of invocation evidence is formed to achieve traceable data retrieval and permission consistency.

Benefits of technology

It improves the accuracy and security of data processing, reduces redundant parsing and vectorization, ensures the traceability of historical states and the stability of permissions, and reduces the risk of unauthorized reuse.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115131A_ABST
    Figure CN123115131A_ABST
Patent Text Reader

Abstract

This invention discloses a method, system, electronic device, and storage medium for constructing, evolving, and traceably retrieving enterprise multi-source data's life memory. It acquires multi-source data generated during enterprise operations, generates source identity identifiers and content fingerprints, and controls idempotent ingestion through a persistent task queue. Separate analysis and memory generation processes form memory units, which are then written with enterprise object identifiers, original evidence pointers, event times, ingestion times, versions, lifecycle states, and permission tags, establishing a memory evolution chain. In response to memory retrieval requests, candidate memory units obtained by fusing keyword indexes, vector indexes, and knowledge graphs are filtered and rearranged according to permission snapshots, query times, and lifecycle states before entering the artificial intelligence model. The generated results are associated with and stored with memory versions, original evidence pointers, and permission snapshots, thereby improving the evolutionary capability of enterprise memory, permission consistency, and traceability of generated results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, enterprise data processing, knowledge engineering, computer security, and information retrieval technology, and in particular to a method, system, electronic device, and storage medium for constructing, evolving, and retrieval the life memory of multi-source enterprise data. Background Technology

[0002] During their operations and management, businesses continuously generate diverse and heterogeneous data, including documents, spreadsheets, images, audio recordings, videos, web pages, meeting minutes, equipment logs, customer records, project documents, and AI interaction records. This data is typically distributed across file systems, business systems, personal terminals, and network services.

[0003] Existing enterprise knowledge bases typically process data through directory management, full-text search, vector databases, knowledge graphs, or search enhancement generation technologies. While these technologies can improve data retrieval and question-answering efficiency, the following technical challenges still exist in long-term enterprise operation scenarios.

[0004] First, raw data, model analysis results, and model-generated content are easily stored together. After multiple updates, the knowledge pages generated by the model may become detached from the original evidence, making it difficult for the system to determine which source, version, and point in time data was used for a particular answer.

[0005] Second, existing knowledge bases often interpret document updates as index updates or content replacements. When the same customer, device, product, project, or system enters different states at different times, the old state may be overwritten, failing to form a continuous memory chain of object evolution, and also failing to restore the valid knowledge state at the time of historical events.

[0006] Third, multiple sources of data may be imported repeatedly. System interruptions, file renaming, directory migration, or rescanning can all cause repeated parsing, vectorization, and generation, resulting in wasted computing resources and multiple conflicting pieces of knowledge content.

[0007] Fourth, keyword retrieval, vector retrieval, and knowledge graph retrieval each have their limitations. Keyword retrieval struggles to handle synonyms, vector retrieval struggles to accurately represent explicit business relationships, and knowledge graph retrieval relies on established entity relationships. A single channel cannot consistently yield results that match the enterprise object, time state, and original evidence.

[0008] Fifth, the positions of company personnel, project member relationships, and confidentiality authorization levels may change. Even if permission checks are performed during retrieval, old caches, old index results, or previously generated contexts may still be reused after permission changes, resulting in unauthorized data entering the AI ​​model context.

[0009] Therefore, there is a need for a technical solution that can transform continuously generated data into a corporate life memory with source, time, version, relationship and permission status, and maintain version validity, permission consistency and evidence traceability during query and artificial intelligence generation processes. Summary of the Invention

[0010] Technical problems to be solved

[0011] The purpose of this invention is to solve the problems arising from the long-term accumulation of multi-source data in enterprises, such as duplicate data ingestion, difficulty in recovering from interruptions, overwriting of historical states, untraceable sources, unstable mixed retrieval, and unauthorized reuse of cached data after permission changes. Technical Solution

[0012] To achieve the above objectives, the first aspect of the present invention provides a method for constructing, evolving, and retrieval traceable memories of multi-source enterprise data, which is executed by a computing device.

[0013] It acquires multi-source data generated during enterprise operation, generates source identity identifiers based on the source type, source location information and the enterprise project to which the multi-source data belongs, and generates content fingerprints based on the content of the multi-source data.

[0014] The data processing task carrying the source identity and content fingerprint is written to the persistent task queue. The data processing task is determined as a skipped task, a new task, an update task, or an invalid task based on the historical content fingerprint corresponding to the source identity.

[0015] The system performs separate analytical processing and memory generation processing on the multi-source data to be processed. The analytical processing generates structured analysis results based on the multi-source data, enterprise object context, and memory pattern constraints; the memory generation processing generates at least one memory unit based on the structured analysis results and the original evidence in the multi-source data.

[0016] Write enterprise object identifiers, original evidence pointers, content fingerprints, independent event times and ingestion times, version identifiers, lifecycle status and permission tags to memory units, and establish evolutionary relationships of supplementation, replacement, conflict or failure based on the content and time relationships between new memory units and existing memory units.

[0017] In response to a memory retrieval request that includes the query content, the access subject, and the query time, candidate memory units are obtained through keyword indexing, vector indexing, and a knowledge graph composed of relationships between memory units, respectively, and the candidate memory units obtained from different retrieval channels are fused together.

[0018] A permission snapshot is generated based on the current permission status of the accessing subject. Before the candidate memory units enter the input context of the artificial intelligence model, the fused candidate memory units are filtered and rearranged using the permission snapshot, query time point, and lifecycle status.

[0019] Memory retrieval results are generated based on the filtered and rearranged candidate memory units. The output fingerprint of the memory retrieval results is calculated and stored in association with the memory unit version used, the original evidence pointer, the permission snapshot identifier, and the retrieval configuration to form a retrieval evidence chain.

[0020] Preferably, the multi-source data includes at least two of the following: documents, tables, images, audio recordings, videos, web page content, business system records, and artificial intelligence interaction records.

[0021] Preferably, the content fingerprint is generated using the SHA-256 algorithm. The source identity can be generated by combining the project identifier, source namespace, source type, normalized path, or business primary key. Separating the source identity from the content fingerprint enables the system to distinguish between "content from the same source changing" and "different sources having the same content".

[0022] Preferably, the persistent task queue stores the task identifier, source identity identifier, content fingerprint, task type, processing stage, task status, retries, and stage checkpoints. Processing stages may include parsing, analysis, memory generation, relation writing, and index updating.

[0023] Preferably, the inputs to the analysis and processing include normalized content, enterprise project objectives, memory patterns, existing memory summaries, and access-permitted associated memories; the outputs include candidate enterprise objects, candidate facts, time information, candidate relationships, conflict information, original evidence locations, and suggested memory types.

[0024] Preferably, the input to the memory generation process includes structured analysis results and original evidence; the output includes memory units, relation edges, manually reviewed items, and index update instructions. The analysis process and the memory generation process each save their processing states, allowing for independent retrying if either stage fails.

[0025] Preferably, the evolutionary relationships include at least addition, supplementation, modification, replacement, conflict, invalidation, and deletion markers. When a new memory unit and an existing memory unit describe the same enterprise object and their contents are mutually exclusive, and the valid status cannot be determined based on event time or source priority, a conflict relationship is established and an audit item is generated.

[0026] Preferably, the system does not physically delete replaced or invalid memory units, but instead preserves their lifecycle state and valid time range. A memory unit only participates in the valid state sorting when the query point falls within its valid time range.

[0027] Preferably, the system is configured with an original evidence layer, an evolutionary memory layer, a rule control layer, and an operation log layer. The original evidence layer uses immutable storage or append-only storage; the evolutionary memory layer stores memory units, relation edges, version information, and vector references; the rule control layer stores memory patterns, model routes, project objectives, and permission policies; and the operation log layer stores ingestion, modification, restoration, retrieval, and generation operations.

[0028] Preferably, keyword retrieval matches the title, body text, tags, and source location information; vector retrieval obtains semantic candidates in the vector index based on the vector representation of the query content; and the knowledge graph performs at least one-hop relation expansion using keyword candidates and vector candidates as seeds.

[0029] Preferably, candidate memory units from different retrieval channels are merged using a reciprocal sorting fusion method, and the candidate scores are adjusted based on at least two of the following: direct relationship, source overlap, common neighbor, object type, version validity, and evidence completeness.

[0030] Preferably, the permission tags include organization tags, job title tags, project tags, data category tags, security classification tags, or access period tags. The permission snapshot records the set of valid permissions for the accessing subject at the time of memory retrieval, along with its snapshot identifier.

[0031] Preferably, the permission snapshot identifier, the version identifier of the candidate memory unit, and the retrieval configuration together constitute the reuse conditions for the memory retrieval cache. When the organization, position, project member relationship, security level authorization, or access period of the access subject changes, a new permission snapshot identifier is generated, and the direct reuse of the candidate context or generation result corresponding to the old permission snapshot is prohibited.

[0032] Preferably, the statements in the output generated by the artificial intelligence model are associated with one or more memory unit identifiers. The system can further parse the memory unit identifiers into the location, content fingerprint, and version information of the original evidence.

[0033] Preferably, when there are unresolved conflicts among candidate memories, the completeness of evidence is below a threshold, or the accessing subject is not authorized to access the necessary evidence, a conflict warning, insufficient evidence status, or rejection result is output, and the unverified model output is not directly written into the evolutionary memory layer.

[0034] The second aspect of the present invention provides an enterprise artificial intelligence life memory system, including a data acquisition and identity module, a task control module, a memory generation module, a memory evolution module, a hybrid retrieval module, an access control module, and a traceability generation module, each module being used to execute the method of the first aspect.

[0035] A third aspect of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the method of the first aspect.

[0036] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method of the first aspect.

[0037] A fifth aspect of the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the method of the first aspect. (Beneficial effects)

[0038] This invention distinguishes data sources from content changes by combining source identity and content fingerprints, and performs idempotent ingestion through persistent queues and phase checkpoints, reducing repeated parsing, repeated vectorization, and repeated model calls, thereby improving recovery capabilities after interruption.

[0039] This invention separates the original evidence, model analysis, and formal memory writing, avoiding the model-generated content from overwriting the original evidence, and improves the verifiability of the knowledge generation results through original evidence pointers and content fingerprints.

[0040] This invention sets event time and ingestion time for memory units and retains historical states through supplementation, substitution, conflict and failure relationships, enabling the system to answer current state questions and historical time point questions respectively.

[0041] This invention combines keyword, vector, and knowledge graph channels to retrieve memories, and integrates version validity and evidence completeness for ranking, which can improve the recall stability of candidate memories when there are inconsistencies in literal expression, semantic expression, and relational expression.

[0042] This invention filters candidate memories using the current permission snapshot before they enter the context of the artificial intelligence model, and incorporates the permission snapshot into the cache reuse conditions, thereby reducing the risk of unauthorized reuse of old contexts or old generated results after permission changes.

[0043] This invention binds the output fingerprint, memory version, original evidence, permission snapshot, and retrieval configuration into a chain of evidence, enabling the generated results to be traced back to the actual enterprise memory and original evidence used. Attached Figure Description

[0044] Figure 1 This is a diagram illustrating the overall architecture of an enterprise AI life memory system provided in an embodiment of the present invention.

[0045] Figure 2 A flowchart illustrating the multi-source data idempotent ingestion process provided in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the memory unit data structure provided in an embodiment of the present invention.

[0047] Figure 4This is a schematic diagram of the memory evolution chain provided in an embodiment of the present invention.

[0048] Figure 5 A flowchart of hybrid memory retrieval provided for embodiments of the present invention.

[0049] Figure 6 A flowchart for generating permission consistency and traceability provided in embodiments of the present invention.

[0050] Figure 7 A flowchart for recording to form corporate life memory provided in an embodiment of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments. The described embodiments are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. ; Embodiment 1: System Overall Architecture

[0052] like Figure 1 As shown, the Enterprise AI Life Memory System 100 includes a multi-source data acquisition module 110, a source identity and content fingerprint module 120, a task control module 130, a two-stage memory generation module 140, a multi-layer memory storage module 150, a hybrid retrieval module 160, an access control module 170, an artificial intelligence model and business intelligence agent module 180, and a call evidence chain module 190.

[0053] The multi-source data acquisition module 110 can acquire enterprise data from local folders, enterprise business interfaces, web clipping portals, manual upload portals, or audio / video processing portals. The data acquisition module does not require different sources to use the same format; instead, it standardizes the data into processable content in subsequent steps.

[0054] The source identity and content fingerprint module 120 generates a source_id based on the project identifier, source namespace, source type, normalized path, or business primary key, and calculates a content_hash for the source content. For example: source_id = Hash(project_id || namespace || source_type || normalized_locator) content_hash = SHA256(normalized_bytes)

[0055] When a file is renamed but the business primary key remains unchanged, the business primary key can be used to maintain the source identity; when the business primary key cannot be obtained, the normalized path can be used as the source location information, and content fingerprints and migration records can be used to assist in identification.

[0056] Task control module 130 writes tasks to a persistent queue. An example of a task record is: task_id, source_id, content_hash, task_type, stage, status, retry_count, checkpoint, created_at.

[0057] Before processing begins, the task control module queries the most recently completed record for the same source_id. If the content_hash is the same, the task is marked as skipped; if they are different, it is marked as updated; if no record is found, it is marked as added; if the source monitoring module detects that the original source has disappeared, it is marked as invalid.

[0058] The two-stage memory generation module 140 includes an analysis submodule 141 and a memory writing submodule 142. The analysis submodule 141 uses a first model call or a first processing procedure to generate structured analysis data; the memory writing submodule 142 uses a second model call or a deterministic processing procedure to generate formal memory units based on the structured analysis data.

[0059] For the artificial intelligence model, the input to the analysis submodule includes at least normalized data, existing enterprise object summaries, and memory patterns; the output adopts a preset structure, for example: { "objects": [], "facts": [], "event_times": [], "relations": [], "conflicts": [], "evidence_spans": []}

[0060] The memory writing submodule 142 generates memory units based on the structured fields described above. The model output must undergo validation for field integrity, evidence location, and permission tags. Validation failures are queued for review and do not become official memories.

[0061] The multi-layered memory storage module 150 includes an original evidence layer 151, an evolutionary memory layer 152, a rule control layer 153, and an operation log layer 154. Each layer is associated using identifiers. The original evidence layer 151 stores imported files or their immutable copies; the evolutionary memory layer 152 stores memory units and relationships; the rule control layer 153 stores enterprise project configurations; and the operation log layer 154 records processing actions over time. Example 2: Enterprise document updates form a memory evolution chain.

[0062] like Figure 2 and Figure 4 As shown, the enterprise project contains the source file "Equipment Maintenance Policy.docx". Upon initial import, it generates a source identity S1 and a content fingerprint H1. The system identifies the enterprise object "Equipment A", the fact "Maintenance cycle is 30 days", and the event time T1 from the document.

[0063] The system generates memory unit M1: memory_id=M1 enterprise_object_id=deviceAnormalized_content=maintenance cycle is 30 days source_id=S1 content_fingerprint=H1 event_time=T1 ingest_time=T2 version_id=V1 lifecycle_state=valid

[0064] The document was subsequently modified to have a "maintenance cycle of 20 days," while the source identity remained S1, and the content fingerprint changed to H2. The task control module detected that H2 differed from the historical H1 and created an update task instead of skipping it.

[0065] The analysis submodule identifies a new fact that describes the same object and attribute as M1, but the event time of the new fact is later than that of M1. The memory writing submodule generates M2, sets version V2, and establishes a substitution relationship between M1 and M2.

[0066] M1 will not be physically deleted; instead, its lifecycle status will be set to "replaced," and its effective end time will be saved. The effective start time for M2 is the effective date of the new system. Users should select M2 when querying "current maintenance cycle"; when querying "maintenance cycle for a historical date," they should select either M1 or M2 based on the query time.

[0067] If another valid equipment manual specifies a maintenance cycle of 25 days, and the system cannot automatically determine the correct value based on time or source priority, then M3 is generated, and a conflict relationship is established between M2 and M3. The system generates an audit item, and before resolving the conflict, it calls the result to mark the conflict, without merging the two values. Example 3: Recording Forms Corporate Life Memory

[0068] like Figure 7 As shown, the user selects a corporate meeting recording file. The system first verifies the file extension and media header, then decodes the audio into waveform data with a uniform sampling rate. Preferably, a 16kHz mono waveform is used as the offline speech recognition input.

[0069] The offline speech-to-text module processes waveform data according to a time window, outputting text segments, start time, and end time. The original recording, waveform fingerprint, transcribed segment, and their time range together constitute the original evidence location information.

[0070] Before the actual recording, the system displays the complete transcript and an AI-organized draft to the user. Users can correct names, numbers, terminology, and speaker information. The revised transcript generates a new content fingerprint, but still maintains a derivational relationship with the original recording source.

[0071] The analysis submodule identifies meeting participants, decision-making items, to-do items, and event time from the reviewed transcript. The memory writing submodule creates meeting memories, decision-making memories, and task memories, and links each memory to the recording time segment and the reviewed text location.

[0072] When a user queries the basis of a decision, the system returns the decision memory and can further locate the corresponding time segment of the original recording, thus enabling the textual answer to be verified by the original audio evidence. Example 4: Hybrid Retrieval and Context Assembly

[0073] like Figure 5 As shown, the system receives query Q, access subject U, query time point QT, and business context BC. The keyword retrieval submodule performs word segmentation retrieval on the title, tags, and body text; the vector retrieval submodule generates query vectors and obtains similarity memories; and the knowledge graph submodule performs relation expansion using two types of candidates as seeds.

[0074] For the i-th candidate memory, the inverse ranking fusion score can be calculated as follows: RRF(i) = Σ 1 / (k + rank_c(i)); where c represents the retrieval channel, rank_c(i) represents the ranking position of the candidate in channel c, and k is a smoothing constant greater than zero.

[0075] Based on the RRF score, we can further calculate: Score(i) = RRF(i) + w1·DirectLink(i) + w2·SourceOverlap(i) + w3·CommonNeighbor(i) + w4·TypeAffinity(i) + w5·VersionValidity(i, QT) + w6·EvidenceCompleteness(i)

[0076] The weights can be configured by the system. VersionValidity is used to determine whether the candidate is valid in QT at the time of query. EvidenceCompleteness is determined based on whether the memory has a readable original source, a valid content fingerprint, and a clear evidence location.

[0077] The system selects top-ranked candidates based on context budget. Context budget can be allocated according to memory text, dialogue history, item rules, and index information. Selected memories are entered into the model context along with their ID, version, and evidence citations. Example 5: Secure Invocation After Employee Permission Changes

[0078] like Figure 6As shown, employee U1 originally belonged to R&D project P1 and had the tags "Organization: R&D Department", "Project: P1", and "Confidentiality Level: Internal". The system generated a permission snapshot PS1 based on the identity directory at that time.

[0079] When U1 initiates a query, a hybrid retrieval process obtains a candidate set C. Before the candidate content enters the AI ​​model context, the access control module filters C using PS1 to obtain an allowed set C1. The system generates an answer A1 and an output fingerprint O1, and saves them as follows: O1 -> {PS1, selected_memory_versions, evidence_refs, retrieval_config}

[0080] Subsequently, U1 exits project P1. The identity directory changes, and the system generates a permission snapshot PS2. Even if the query content is the same as before, because PS2 is different from PS1, the system does not directly reuse the candidate context or answer cache generated based on PS1.

[0081] The system re-executes permission filtering, and memory units that were previously only open to member P1 will not enter the model context. This avoids continuing to obtain project memories through the old cache after permission is revoked. Example Six: Traceable Generation and Memory Writeback

[0082] The AI ​​model generates answers based on a memory context filtered by permissions. The system saves the model configuration, retrieval configuration, selected memory version, and original evidence pointers, and calculates the output fingerprint for the final output.

[0083] When a user requests that the generated results be saved as enterprise knowledge, the system does not directly overwrite existing memories. Instead, it treats the generated results as candidate memory events, records the memory versions and evidence they reference, and performs rule verification or manual review.

[0084] After approval, the system generates a new memory unit and version identifier for the candidate content; if the review fails, the review record is retained but not added to the formal search index. Example 7: Electronic Devices

[0085] Electronic devices include one or more processors, memory, network interfaces, and input / output interfaces. The memory stores computer programs and stored data. When the processor executes the computer program, it performs multi-source data acquisition, fingerprint calculation, task control, model invocation, memory evolution, hybrid retrieval, access control filtering, and evidence chain writing.

[0086] The electronic device can be an enterprise server, a personal computer, an edge computing device, or a computing cluster consisting of multiple nodes. Different modules can run on the same device or on different devices in a controlled network.

[0087] Those skilled in the art will understand that modifications or equivalent substitutions can be made to the above embodiments without departing from the concept of the present invention. All modifications, equivalent substitutions, and improvements within the scope of the claims of this invention should be included within the protection scope of this invention.

Claims

1. A method for constructing, evolving, and retrieval traceable memories of multi-source enterprise data, characterized in that, Performed by a computing device, the method includes: Acquire multi-source data generated during enterprise operation, generate source identity identifiers based on the source type, source location information and enterprise project to which the multi-source data belongs, and generate content fingerprints based on the content of the multi-source data; The data processing task carrying the source identity and the content fingerprint is written into the persistent task queue, and the data processing task is determined as a skipped task, a new task, an update task or an invalid task according to the historical content fingerprint corresponding to the source identity. The analysis and memory generation processes are performed separately on the multi-source data to be processed. The analysis process generates structured analysis results based on the multi-source data, enterprise object context, and memory pattern constraints. The memory generation process generates at least one memory unit based on the structured analysis results and the original evidence in the multi-source data. Write the enterprise object identifier, original evidence pointer, content fingerprint, mutually independent event time and ingestion time, version identifier, lifecycle status and permission tag into the memory unit, and establish the evolutionary relationship of supplementation, substitution, conflict or failure according to the content relationship and time relationship between the memory unit and the existing memory unit to form the memory evolution chain corresponding to the enterprise object, and retain the existing memory units before the establishment of the evolutionary relationship; In response to a memory retrieval request containing query content, access subject, and query time, candidate memory units are obtained through keyword indexing, vector indexing, and a knowledge graph composed of relationships between memory units, respectively, and candidate memory units obtained from different retrieval channels are fused. A permission snapshot is generated based on the current permission status of the accessing subject. Before the candidate memory units enter the input context of the artificial intelligence model, the fused candidate memory units are filtered and rearranged using the permission snapshot, the query time point, and the lifecycle status. Memory retrieval results are generated based on the filtered and rearranged candidate memory units. The output fingerprint of the memory retrieval results is calculated, and the output fingerprint is associated with and stored with the memory unit version, the original evidence pointer, the permission snapshot identifier, and the retrieval configuration to form a retrieval evidence chain that can be reversed from the memory retrieval results to the original evidence. The permission snapshot identifier, the memory unit version, and the retrieval configuration together constitute the reuse conditions of the memory retrieval cache.

2. The method according to claim 1, characterized in that, The multi-source data includes at least two of the following: documents, tables, images, audio recordings, videos, web page content, business system records, and AI interaction records. For audio recordings or videos, media format standardization and speech transcription are first performed on the local computing device to generate transcribed text with time segments, and the original media file, the transcribed text, and the manual revision results are associated with the same source identity.

3. The method according to claim 1, characterized in that, The persistent task queue stores the task identifier, source identity identifier, content fingerprint, processing stage, task status, number of retries, and stage checkpoint for each data processing task; after an abnormal termination of a task, processing is resumed from the most recently completed stage checkpoint.

4. The method according to claim 1, characterized in that, The structured analysis results include at least candidate enterprise objects, candidate facts, time information, candidate relationships, conflict information, original evidence locations, and suggested memory types; the memory generation process also generates relationship edges, manually reviewed items, and index update instructions.

5. The method according to claim 1, characterized in that, The evolutionary relationship includes at least one of the following: addition, supplementation, modification, replacement, conflict, invalidation, and deletion markers; when a new memory unit and an existing memory unit describe the same enterprise object and their contents are mutually exclusive, and the valid status cannot be automatically determined based on event time or source priority, a conflict relationship is established and an audit item awaiting manual processing is generated.

6. The method according to claim 1, characterized in that, The original evidence, memory units, and rule configurations are stored in the original evidence layer, the evolutionary memory layer, and the rule control layer, respectively. The original evidence layer uses immutable storage or only append-only storage, and the content generated by the artificial intelligence model is not overwritten into the original evidence layer.

7. The method according to claim 1, characterized in that, The fusion of candidate memory units obtained from different retrieval channels includes: performing inverse sorting fusion based on the ranking position of the candidate memory units in each retrieval channel, and weighting the fusion result according to at least two of the following: direct relationship, source overlap, common neighbor, object type, version validity, and evidence completeness.

8. The method according to claim 1, characterized in that, The relation edges of the knowledge graph are generated based on at least one of the following: explicit links between memory units, common original sources, common neighbors, and object type affinity; at least one-hop relation expansion is performed using candidate memory units obtained from keyword retrieval and vector retrieval as seeds.

9. The method according to claim 1, characterized in that, The permission tags include at least one of the following: organization tag, position tag, project tag, data category tag, security level tag, and access period tag; in response to a change in at least one of the following: organization, position, project member relationship, security level authorization, or access period of the access subject, a new permission snapshot identifier is generated, and the reuse of candidate memory contexts or memory retrieval results associated with the old permission snapshot identifier is prohibited.

10. The method according to claim 1, characterized in that, When there are unresolved conflicts, the integrity of the original evidence is below the threshold, or the accessing subject is not authorized to access the necessary evidence after filtering and rearranging the candidate memory units, a conflict prompt, insufficient evidence status, or rejection result is generated; the output generated by the artificial intelligence model is only written into the memory evolution chain as a new memory event after passing the rule verification or manual review.

11. A corporate artificial intelligence life memory system, characterized in that, include: The data acquisition and identity module is used to acquire multi-source data from enterprises and generate source identity identifiers and content fingerprints. The task control module is used to write data processing tasks into a persistent task queue and control idempotent ingestion and anomaly recovery based on historical content fingerprints. The memory generation module is used to generate memory units through separate analysis and memory generation processes. The memory evolution module is used to write enterprise object identifiers, original evidence pointers, time information, version information, lifecycle status and permission tags into memory units, and to establish a memory evolution chain; A hybrid retrieval module is used to obtain and fuse candidate memory units through keyword indexing, vector indexing, and knowledge graphs; The access control module is used to generate an access snapshot based on the current access status of the access subject and to filter access permissions before candidate memory units enter the input context of the artificial intelligence model. The traceable generation module is used to generate memory retrieval results based on the filtered candidate memory units, and to associate and store the memory retrieval results with the memory unit version, the original evidence pointer, and the permission snapshot identifier.

12. The system according to claim 11, characterized in that, It also includes a local audio and video processing module, which is used to perform audio and video decoding, speech transcription and manual review of transcription results on enterprise terminal devices, and submit the reviewed transcription results as new multi-source data to the task control module.

13. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the electronic device to perform the method according to any one of claims 1 to 10.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 10.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 10.