A hierarchical memory and multi-tenant knowledge collaboration method, storage medium, and computer program product for a large-scale model intelligent agent system in a financial fixed-income trading scenario.

By employing a hierarchical memory and multi-tenant knowledge collaboration approach, the problems of contextual noise and multi-tenant isolation in large language models under financial fixed-income trading scenarios are solved. This enables structured review and risk control of transaction facts, reduces the risk of secondary leakage, and improves the flexibility and traceability of business expansion.

CN123115318APending Publication Date: 2026-11-17BEIJING KUAQUO INFORMATION TECH CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611229983.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the context of financial fixed-income trading, large language models have limited and costly single-time contexts, are difficult to reuse across rounds and sessions, have high requirements for multi-tenant isolation, and lack knowledge binding with access control, which poses a risk of unauthorized injection. Transaction facts are difficult to review and audit in a structured manner, and toolchain risks are difficult to control.

Method used

By adopting a hierarchical memory and multi-tenant knowledge collaboration approach, the system rules area, session memory area, task status area, and retrieval evidence area are assembled in partitions by maintaining short-term, medium-term, and long-term memory data. This generates a change set data structure, performs access control and outbound governance, ensures consistency between knowledge inbound and outbound governance, and reduces contextual noise and risks.

Benefits of technology

It achieves tenant and team isolation and traceability, structured and auditable transaction facts, declarative evolution of extraction rules, reduces the risk of secondary leakage, and improves the flexibility and traceability of business expansion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115318A_ABST
    Figure CN123115318A_ABST
Patent Text Reader

Abstract

This invention discloses a hierarchical memory and multi-tenant knowledge collaboration method for a large-scale intelligent agent system in a financial fixed-income trading scenario. The method involves injecting short-term, medium-term, and long-term memory contexts into system partitions, generating change sets and compressed anchors; controlling retrieval boundaries using multi-tenant knowledge and session snapshots; governance and isomorphic pre-outbound scanning during the synthesis of transaction parameters; and structurally storing and recording skill declaration field schemas after restricted generation, verification, and optional human-machine confirmation, along with the skill version. It also supports the sequential merging of system default strategies, skill strategies, and session coverage strategies to achieve scenario-based customization of memory behavior. The advantages of this invention include reduced contextual noise, traceable isolation between tenants and teams, structured and auditable transaction facts, declarative evolution of extraction rules with version attribution, consistency between knowledge inbound and outbound governance, and reduced secondary leakage.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent agent technology applicable to finance, and particularly to a hierarchical memory and multi-tenant knowledge collaboration method, storage medium, and computer program product for large-scale intelligent agent systems in financial fixed-income trading scenarios. Background Technology

[0002] In fixed income trading and investment research support scenarios, in addition to dialogue, intelligent agents also need to be reused across rounds and sessions: (1) key communication points between traders and clients; (2) private domain knowledge such as internal research reports or rating notes; (3) structured elements such as inquiries, quotations, and receipts. Large language models have limited context and high costs per instance. Simply stacking the entire text will not only exceed the budget but also easily cause fact drift. At the same time, institutions have high requirements for isolation of multiple tenants and multiple teams. If knowledge is not bound to access control (ABAC / RBAC) and session visibility, there is a risk of unauthorized injection. If the trading side only has unstructured dialogue records, it is difficult to review, audit, and align with risk control rules.

[0003] The existing technology is: (1) General RAG framework: including document slicing, vector retrieval, and prompt filling, often lacks deep coupling with outbound tool governance and in-session visibility snapshot; (2) Single “memory bar”: Memories are stored in the form of key-value pairs or summaries, but they are often not hierarchical and do not have a structured change set and user error correction loop; (3) Handwritten ETL (Extract, Transform, Load): The table structure and extraction script are fixed for each business, which is costly and inconvenient to evolve with the business capability version. It is also difficult to declare the extraction rules on the skill profile and make the runtime uniformly execute the gate.

[0004] The disadvantages of existing technology are: (1) Lack of hierarchical structure: Short-term context, cross-session mid-term working set and long-term hierarchical knowledge are mixed together, making it impossible to compress and retrieve on demand, and also making it difficult to present the "memory change list" in rounds; (2) Toolchain risk: When retrieving fragments are incorporated into transaction / external tool parameters, if there is no governance consistent with the pre-exit scan of the same structure, residual sensitive information may be transmitted. (3) The transaction facts are not verifiable: Natural language archiving alone is insufficient to support threshold rules, playback and auditing; (4) Business expansion rigidity: If each type of security or asset segment process has its own extraction logic, it will be redundant and difficult to perform regression testing and version attribution.

[0005] CN121501919A discloses an automatic update method for an intelligent customer service knowledge base based on a large language model, emphasizing multi-source document collection, candidate item extraction, patch merging, and effective switching. Its advantage is automated knowledge updates, but it leans towards a closed loop for customer service knowledge domain and item updates.

[0006] IN202511113728A discloses a computer implementation system for managing context exchange in machine learning workloads, emphasizing scope authorization, capsule representation, visibility, and governance records. Its strength lies in its governance track design, but it leans towards a general context governance and cross-model exchange framework.

[0007] The technical solutions in the two documents mentioned above have at least the following shortcomings: (1) Insufficient hierarchical memory collaboration: CN121501919A is biased towards candidate item updates, and IN202511113728A is biased towards governance track. Neither of them fully solves the integrated mechanism of "short-term / medium-term / long-term three-layer linkage, context partitioning assembly, and budget compression to protect anchor points"; (2) Insufficient granularity of multi-tenant business collaboration: Although the concepts of scope or permission are involved, the implementation of “organizational identifier, team scope, and session visibility snapshot” to suppress the expansion of the search scope is not complete. (3) Insufficient binding of transaction domain skills: None of them have formed an end-to-end process of "fixed income skill profile declaration field schema, extraction trigger, constraint generation, rule verification, and versioning and storage"; (4) Insufficient coupling of external parameter link governance: The consistency scanning mechanism before and after the retrieval evidence enters the transaction / external tool parameters is mostly described in general in the comparison scheme, lacking fine-grained coupling with the fixed income transaction parameter synthesis link; (5) Insufficient memory detail and weak scenario customization capability: The comparison solutions are mostly centered on general knowledge updates or context governance, without detailing the differences in compression, structured extraction, and role template dimension accumulation at each level of the short-term / medium-term / long-term, and without forming a closed-loop capability of "customizing memory strategies based on business skills according to specific business scenarios".

[0008] The above are the areas that need to be improved in this application. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a hierarchical memory and multi-tenant knowledge collaboration method, storage medium and computer program product for a large model intelligent agent system in the context of financial fixed income trading, which reduces context noise, isolates and traces tenants and teams, structures and audits transaction facts, extracts rules in a declarative evolution and version attribution, ensures consistency between knowledge inbound and outbound governance, and reduces secondary leakage.

[0010] To address the above technical issues, this invention provides a hierarchical memory and multi-tenant knowledge collaboration method for large-model intelligent agent systems in financial fixed-income trading scenarios, comprising the following steps: Maintain at least two of the following: short-term memory data, tenant-internal medium-term memory data, and long-term hierarchical knowledge base data. When providing the generative context to the large model intelligent agent system, partition and assemble the system rule area, session memory area, task state area, retrieval evidence area, and tool structured receipt area, and prohibit disordered stacking. In response to a session round or selected transaction boundary, generate and output a changeset data structure that describes the categories of added, updated, or deleted memory entries; Trigger summary-style compression of short-term memory data when the context budget threshold is met, and retain anchor references for key decision entries; and Knowledge base search results associated with organizational identity and user team scope are recalled after filtering by access control attributes for use in the search evidence area; The method further includes: defining at least one skill profile declaration field schema associated with fixed-income transactions; inputting at least one of user session text data and tool structured output data into a constrained generation framework under the extraction trigger conditions defined by the skill profile to produce an intermediate fact data structure that conforms to the field schema; and persisting the intermediate fact data structure as at least one of medium-term memory data or long-term knowledge base data after the intermediate fact data structure passes the validation of the pattern and business rules and passes the optional human-machine confirmation, and identifying the skill version of the persisted action record.

[0011] The access control writes a knowledge visibility snapshot during the duration of the session to suppress the expansion of the retrieval scope.

[0012] When the text output from the retrieval evidence area enters the synthesis path that will generate the call parameters for transactions or external tools, a policy scan of the same class as the outbound governance is applied before the outbound parameters are executed or the tool is executed.

[0013] The summary-type compression is preferentially applied to short-term session memory in order to maintain a balance between timeliness and cost.

[0014] The extraction triggering conditions include at least one of the tool call completion event or the human-machine adjudication completion event.

[0015] This also includes keeping memory entries that do not meet the approval criteria in an isolated candidate state until they are approved by governance and then transferred to a long-term knowledge base index that can be retrieved.

[0016] The memory strategy is obtained by sequentially merging the default configuration of the large model intelligent agent system, the skill memory configuration memory_profile, and the session-level memory configuration overridden memory_profile_override. The session-level memory configuration overridden has a higher priority than the skill memory configuration, and the skill memory configuration has a higher priority than the default configuration of the large model intelligent agent system. This allows for scenario-based customization of at least one of the following: compressed window, recall scope, retrieval method, and privacy level, thereby improving traceability and consistency of risk control.

[0017] In addition, the present invention provides a storage medium, which is a computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the steps of the hierarchical memory and multi-tenant knowledge collaboration method described above.

[0018] In addition, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the hierarchical memory and multi-tenant knowledge collaboration method as described above.

[0019] The technical solution proposed in this invention has at least the following beneficial effects: (1) By hierarchically managing short-term memory, medium-term memory and long-term knowledge base, and assembling context according to system rules area, conversation memory area, task status area, retrieval evidence area and tool structured receipt area, the risk of context noise and unreliable trading advice is reduced; (2) By filtering the search results based on the organization identifier, team scope, entry confidentiality level and session knowledge visibility snapshot, the tenant and team isolation is made traceable; (3) After the tool call success event is triggered, the transaction facts are extracted in a constrained manner, the pattern is validated, the business rules are validated and the user is confirmed according to the field pattern, and the source receipt and anchor point are associated to update the intermediate memory, so that the transaction facts are structured and auditable. (4) By recording skill version identifiers, extractor version identifiers, source anchors and approval records, the extraction rules can evolve declaratively and be attributed to versions; (5) By linking the desensitization and classification when knowledge is entered into the database, the allowed block filtering during retrieval, and the pre-inspection before the synthesis of tool parameters, the knowledge utilization link is kept consistent with the outbound governance, thereby reducing the risk of secondary leakage caused by retrieval evidence. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is an overall flowchart of a specific embodiment of the present invention; Figure 2 This is a schematic diagram of hierarchical memory and context partitioning injection in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the change set, compression, and anchor point in an embodiment of the present invention; Figure 4 This is a schematic diagram of the business extraction-verification-database entry pipeline according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the merging of business memory strategies and scenario customization in an embodiment of the present invention. Detailed Implementation

[0021] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The executing entity of a specific embodiment of the present invention is a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, mobile phone, etc., or an electronic device that implements the above functions.

[0022] like Figure 1 This invention provides a hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios, comprising the following steps: Maintain at least two of the following: short-term memory data, tenant-internal medium-term memory data, and long-term hierarchical knowledge base data. When providing the generative context to the large model intelligent agent system, partition and assemble the system rule area, session memory area, task state area, retrieval evidence area, and tool structured receipt area, and prohibit disordered stacking. like Figure 2 As shown, short-term session memory, medium-term in-tenant working set memory, and long-term hierarchical knowledge base are maintained, and inter-layer boundaries and read / write strategies are established; the memory levels in Table 1 correspond to the hierarchical responsibility boundaries: used to define the differences between short, medium, and long-term memory in terms of tenants and retrieval scope.

[0023] Table 1: Memory Hierarchy ; Table 2: Context Partitioning .

[0024] Context partitioning and assembly: When generating the context, the system rules area, session memory area, task status area, retrieval evidence area and tool receipt area are partitioned and assembled separately, and unordered stack mixing is prohibited.

[0025] like Figure 3 As shown, in response to session rounds or selected transaction boundaries, a changeset data structure describing the categories of added, updated, or deleted memory entries is generated and output. Trigger summary-style compression of short-term memory data when the context budget threshold is met, and retain anchor references for key decision entries; and Knowledge base search results associated with organizational identity and user team scope are recalled after filtering by access control attributes for use in the search evidence area; like Figure 4 As shown, it further includes: enabling at least one skill profile declaration field schema associated with fixed income transactions; under the extraction trigger conditions defined by the skill profile, inputting at least one of user session text data and tool structured output data into a constrained generation framework to produce an intermediate fact data structure that conforms to the field schema, and supporting the loading of different skill memory templates according to specific scenarios such as quotations, inquiries, transaction reviews, and counterparty profiles.

[0026] The intermediate fact data structure undergoes pattern validation and business rule validation before persistence. The business rule validation includes two methods: (1) Automatic rule verification method: The rule engine automatically verifies the validity period of the quote, the price or yield range, the bond identifier format, the consistency of the transaction direction, and the consistency of the receipt handle association; (2) Review and verification method: When high-risk, low-confidence or rule conflict conditions are triggered, the manual confirmation or second-round decision process is initiated before it is determined whether to release the application.

[0027] After the intermediate fact data structure passes pattern validation, business rule validation, and optional human-machine confirmation, it is persisted according to risk level and stability: low-risk facts that can be directly reused in the current session are written into intermediate memory data; facts that meet the conditions for long-term retention (including cross-session stability, approval, and permission matching) are written into long-term knowledge base data; facts that fail validation or confirmation are not persisted. The system records skill version identifiers for persistent actions; long-term memory entries are structured and archived according to role template dimensions, which include transaction roles, risk preferences, communication preferences, and authorization boundaries.

[0028] The access control writes a knowledge visibility snapshot during the duration of the session to suppress the expansion of the retrieval scope.

[0029] When the text output from the retrieval evidence area enters the synthesis path that will generate the call parameters for transactions or external tools, a policy scan of the same class as the outbound governance is applied before the outbound parameters are executed or the tool is executed.

[0030] The summary-type compression is preferentially applied to short-term session memory in order to maintain a balance between timeliness and cost.

[0031] The extraction triggering conditions include at least one of the tool call completion event or the human-machine adjudication completion event.

[0032] This also includes keeping memory entries that do not meet the approval criteria in an isolated candidate state until they are approved by governance and then transferred to a long-term knowledge base index that can be retrieved.

[0033] like Figure 5 As shown, the memory strategy is obtained by sequentially merging the default configuration of the large model intelligent agent system, the skill memory configuration memory_profile, and the session-level memory configuration overridden memory_profile_override. Among them, the session-level memory configuration overridden has a higher priority than the skill memory configuration, and the skill memory configuration has a higher priority than the default configuration of the large model intelligent agent system. This allows for scenario-based customization of at least one of the following: compressed window, recall scope, retrieval method, and privacy level, thereby improving traceability and consistency of risk control.

[0034] Table 3: Boundaries of Three-Layer Memory Refinement ; Table 3 illustrates the differentiated realization and division of roles in the short, medium, and long term. Specific Implementation Example 1: This embodiment uses the internal rating Wiki of an institution to assist in the pricing of spot bonds as an example to illustrate the controlled entry, retrieval, and governance process of the evidence-to-tool parameter path of a long-term hierarchical knowledge base.

[0036] First, the system receives internal rating Wiki documents and assigns the document the associated organization identifier, the team to which it belongs, the document's security level, permitted business scenarios, and the source version identifier. The system then divides the document into multiple knowledge blocks according to preset segmentation rules. Each knowledge block retains the original text's location anchor, source version identifier, and permission attributes, ensuring that subsequent content retrieval traces back to the specific document location, rather than simply retaining a summary text detached from the source.

[0037] Secondly, the system performs anonymization and hierarchical processing on each knowledge block. For trader names, client names, account identifiers, contact information, or other information that should not be included in the retrieval evidence area, it is replaced with preset aliases, masks, or general descriptions according to anonymization rules. Highly sensitive information that cannot be processed by anonymization rules is either not written to the retrieval index or written to a hierarchical storage area visible only to specific roles. The processed knowledge blocks, along with the organization identifier, team scope, security classification, original text anchor, and document version, are written to the long-term hierarchical knowledge base. Through this step, long-term knowledge accumulation does not involve indiscriminately throwing Wiki original text into the vector library, but rather forming executable retrieval boundaries during the data entry stage, reducing the risk of cross-team knowledge being mixed into the context.

[0038] When a user initiates a bond price inquiry task, the system parses the task status from the user's input, including bond identifier, buy / sell direction, target size, term, or price constraints, and generates a knowledge visibility snapshot during the current session. This snapshot at least fixes the institution identifier, user team scope, role permissions, and accessible knowledge security level; within the same session, subsequent searches are filtered based on this snapshot, and the search scope is not expanded due to model-generated content or temporary keyword expansion.

[0039] The system retrieves candidate knowledge blocks from the long-term knowledge base based on current bond pricing tasks. It first filters these blocks using institutional identifiers, team scope, security level, and session knowledge visibility snapshots. Then, it sorts the filtered knowledge blocks according to relevance, source version, and business timeliness. Only blocks are allowed to be written to the retrieval evidence area; knowledge blocks that fail the access control filtering are not included in the large model context or participate in subsequent tool parameter synthesis. This isolates tenants and teams, providing replayable and traceable snapshot evidence, and reduces contextual noise and unreliable trading advice risks caused by irrelevant rating information entering the context.

[0040] The system will write allowed blocks, task states, and session memories into the retrieval evidence area, task state area, and session memory area respectively, and construct the query tool draft according to the partition order; the original full text of the Wiki, unfiltered candidate blocks, or tool receipts must not be directly mixed into the tool parameters. If the retrieval evidence is used to generate call parameters for transactions or external tools, a pre-check will be performed before the draft enters the actual tool execution path. The pre-check will at least check field integrity, knowledge block permission attributes, residual sensitive information, consistency between voucher identifiers and task states, and the existence of reference anchors; if the pre-check fails, the draft will be prevented from entering the outbound or tool execution path and the corresponding reason will be output. By linking knowledge entry, retrieval filtering, and tool draft pre-check, the knowledge utilization chain and outbound governance adopt a consistent control approach, reducing the risk of secondary leakage caused by retrieval evidence. Specific Implementation Example 2: like Figure 4 As shown in the example, this embodiment uses a multi-round spot bond market-making session to illustrate the process of the tool completing the structured business fact extraction, medium-term memory update, and short-term memory compression triggered by the event.

[0042] In multi-round market-making sessions, the system stores user input, model output, task status, and tool structured receipts in the session memory area, task status area, and tool structured receipt area, respectively. The task status includes at least the security identifier, buy / sell direction, quantity range, current quote or yield, quote validity period, and counterparty alias; the tool structured receipt includes at least the tool call result status, receipt handle, and call time. The system does not directly use all historical dialogues as prompts for subsequent rounds; instead, it selects short-term memory fragments, task statuses, and tool receipts according to preset partitioning and budget strategies for assembly.

[0043] When the quoting or inquiry tool returns a success status (tool_success), the extraction trigger condition related to the current bond market scenario in the skill profile is met. The system extracts intermediate fact data structures from the current session text and the successful tool receipt based on the schema fields declared in the corresponding skill profile. These fields may include bond identifier, trading direction, size range, price or yield quote, counterparty alias, quote validity period, receipt handle, and extraction confidence level. The constrained generation framework only outputs candidate data that conforms to the schema fields and does not directly write free text into the intermediate table.

[0044] The system performs schema validation and business rule validation on the intermediate fact data structure. Schema validation verifies field types, required fields, and enumerated values; business rule validation verifies the consistency between the quote validity period, price or yield range, bond identifier format, transaction direction, and receipt handle with the current tool call. If the validation fails, the candidate fact is retained as a record to be corrected, and the intermediate memory is not updated; if the validation passes, the system displays the candidate fact as a user confirmation card, which includes at least the field to be updated, the old and new values, the source receipt, and the extraction credibility.

[0045] After the user confirms the addition or update of content in the card, the system determines the scope of the intermediate memory based on the session identifier, task identifier, voucher identifier, and counterparty alias, and performs an upsert operation on the intermediate table; simultaneously, it records the skill version identifier, extractor version identifier, confirmation time, and source anchor. If the user rejects or corrects a candidate fact, the system generates a corresponding deletion or update changeset, instead of using unconfirmed content as stable facts for subsequent sessions. Therefore, the generation, confirmation, and updating of transaction facts can all be linked to specific sessions and tool receipts, facilitating auditing and review.

[0046] When the context budget of short-term session memory reaches a threshold, the system prioritizes summary compression of the short-term session memory. During compression, anchor references related to confirmed quotes, user corrections, tool success receipts, and interim table update records are preserved; subsequent rounds locate the original session fragments or tool receipts through anchor references, without continuously carrying all historical text. This maintains the continuity of multi-round market making while reducing the consumption of context budget by redundant historical content and mitigating the risk of generating unreliable trading advice due to context noise. Specific Implementation Example 3: like Figure 3 and Figure 4 As shown, this embodiment takes writing transaction preferences or counterparty profiles accumulated across sessions into a long-term knowledge base as an example to illustrate the processing of asynchronous candidate entries, approval or timeout failure, long-term index promotion, and high-risk secondary adjudication.

[0048] When a session round ends, a tool call succeeds, or a user completes a decision, the system, based on the trigger conditions of the corresponding skill profile, performs constrained generation on the session text, tool structured receipts, or confirmed interim facts to form asynchronous candidate entries. Candidate entries must at least include structured fields, the affiliated organization identifier, team scope, source session or receipt anchor, skill version identifier, extraction credibility, risk level, and candidate generation time. Candidate entries are initially written to an isolated candidate area and are not directly entered into the searchable long-term knowledge base index.

[0049] For low-risk candidate entries, the system initiates an approval process after the candidate is generated. The approval process can be completed by users with the appropriate role permissions or by preset governance rules. During approval, the system verifies the candidate entry's source anchor, field schema, business rules, tenant affiliation, and team scope. After verification and approval, the system promotes the candidate entry from the isolated candidate area to a formal entry in the long-term hierarchical knowledge base index, and retains the candidate identifier, approval record, and skill version identifier to ensure the traceability of the long-term entry's generation path.

[0050] The system configures a Time-To-Live (TTL) for unapproved candidate entries. If a candidate entry fails to pass approval within the TTL period, or is withdrawn by the user or rejected by business rules before approval, the candidate entry is marked as invalid and removed from the isolated candidate area. This candidate entry cannot be retrieved from the long-term knowledge base, nor can it be written into the parameter synthesis path of subsequent transaction tools. This approval and TTL invalidation mechanism prevents the long-term accumulation of short-term, occasional, or unconfirmed session facts.

[0051] For candidate entries whose risk level reaches a preset high-risk threshold, the system still performs a secondary review after the general approval process. The secondary review at least verifies whether the candidate content contains sensitive identity information, whether it exceeds the visibility range of the original session, whether it may change the counterparty's authorization boundaries or risk preferences, and whether the access security level of long-term entries should be increased. If the secondary review fails, the candidate entry remains isolated or invalidated; only after the secondary review passes is it allowed to be written into the long-term index with the corresponding security level.

[0052] When long-term entries are retrieved in subsequent sessions after promotion, they still need to be filtered based on organizational identifier, team scope, entry confidentiality level, and session knowledge visibility snapshot. Thus, candidate generation, approval, invalidation, secondary adjudication, promotion, and subsequent retrieval form a closed loop: it supports the sedimentation of stable business facts into reusable knowledge, and ensures that the writing of long-term knowledge is consistent with access control, risk governance, and audit traceability.

[0053] The present invention also provides a computer-readable storage medium having computer-readable program instructions stored thereon, the computer-readable program instructions being used to execute the hierarchical memory and multi-tenant knowledge collaboration method in the above embodiments.

[0054] The computer-readable storage medium provided in this application is, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium is any tangible medium that contains or stores a program that is used by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium is transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0055] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the hierarchical memory and multi-tenant knowledge collaboration method as described above.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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 hierarchical memory and multi-tenant knowledge collaboration method for a large-scale model intelligent agent system in a financial fixed-income trading scenario, characterized in that: The steps include the following: Maintain at least two of the following: short-term memory data, tenant-intermediate memory data, and long-term hierarchical knowledge base data. When providing generative context to large model intelligent agent systems, partition and assemble the system rule area, session memory area, task state area, retrieval evidence area, and tool structured receipt area. In response to a session round or selected transaction boundary, generate and output a changeset data structure that describes the categories of added, updated, or deleted memory entries; When the context budget threshold is met, a summary-type compression of short-term memory data is triggered, and anchor references are preserved for key decision entries; as well as Knowledge base search results associated with organizational identity and user team scope are recalled after filtering by access control attributes for use in the search evidence area; The method further includes: defining at least one skill profile declaration field schema associated with fixed-income transactions; inputting at least one of user session text data and tool structured output data into a constrained generation framework under the extraction trigger conditions defined by the skill profile to produce an intermediate fact data structure that conforms to the field schema; and persisting the intermediate fact data structure as at least one of medium-term memory data or long-term knowledge base data after the intermediate fact data structure passes the validation of the pattern and business rules and passes the optional human-machine confirmation, and identifying the skill version of the persisted action record.

2. The hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios according to claim 1, characterized in that: The access control writes a knowledge visibility snapshot during the duration of the session to suppress the expansion of the retrieval scope.

3. The hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios according to claim 1, characterized in that: When the text output from the retrieval evidence area enters the synthesis path that will generate the call parameters for transactions or external tools, a policy scan of the same class as the outbound governance is applied before the outbound parameters are executed or the tool is executed.

4. The hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios according to claim 1, characterized in that: The summary-type compression is preferentially applied to short-term session memory in order to maintain a balance between timeliness and cost.

5. The hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios according to claim 1, characterized in that: The extraction triggering conditions include at least one of the tool call completion event or the human-machine adjudication completion event.

6. The hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios according to claim 1, characterized in that: This also includes keeping memory entries that do not meet the approval criteria in an isolated candidate state until they are approved by governance and then transferred to a long-term knowledge base index that can be retrieved.

7. The hierarchical memory and multi-tenant knowledge collaboration method for large-scale model intelligent agent systems in financial fixed-income trading scenarios according to claim 1, characterized in that: The memory strategy is obtained by sequentially merging the default configuration of the large model intelligent agent system, the skill memory configuration, and the conversation-level memory configuration coverage to obtain an effective strategy. Among them, the conversation-level memory configuration coverage has a higher priority than the skill memory configuration, and the skill memory configuration has a higher priority than the default configuration of the large model intelligent agent system, so as to customize at least one of the following in a scenario: compressed window, recall scope, retrieval method, and privacy level.

8. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the hierarchical memory and multi-tenant knowledge collaboration method as described in any one of claims 1 to 7.

9. A computer program product, characterized in that: It includes a computer program that, when executed by a processor, implements the steps of the hierarchical memory and multi-tenant knowledge collaboration method as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Intelligent customer service knowledge base automatic updating method based on large language model

    CN121501919A