An ai agent memory integration and consolidation method and system of a bionic sleep mechanism

Abstract

The application discloses an AI intelligent body memory integration and consolidation method and system of a bionic sleep mechanism, and relates to the technical field of artificial intelligence memory management. The method comprises the following steps: automatically triggering a memory integration cycle during a system low-load period; scanning newly added memories of the day, discovering cross-memory semantic associations through a graph traversal algorithm and a dopamine gating mechanism, and merging the cross-memory semantic associations into memory clusters; removing repeated information; extracting abstract rules from specific events; and performing adaptive forgetting based on value scoring. The application simulates the memory consolidation process in human sleep, realizes the association integration, de-duplication and abstraction of memories, and intelligent forgetting, and improves the memory quality and retrieval efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence memory management technology, specifically to a biomimetic method and system that simulates the memory consolidation mechanism during human sleep, and performs association discovery, deduplication integration, and abstraction processing on the memories accumulated by an AI agent during the day. Background Technology

[0002] Existing AI agents accumulate a large amount of memory data during operation, including dialogue interaction records, tool call results, user preference information, etc. These memories are usually stored in the form of raw logs, which have the following problems: fragmented memory, with information on the same topic scattered across multiple records and unable to be automatically linked; high redundancy, with the same facts being recorded repeatedly, consuming a lot of storage space; lack of abstraction, making it impossible to extract reusable rules and patterns from specific events; and a single memory forgetting strategy, with simple time decay mechanisms unable to distinguish between high-value and low-value memories.

[0003] Current memory management technologies mainly employ full storage of vector databases, time-based expiration deletion, or capacity-based LRU eviction strategies. These methods lack assessment and structured integration of memory quality and cannot simulate the cognitive processes of "memory replay, association integration, and elimination of useless information" during human sleep. Summary of the Invention

[0004] This invention aims to provide a memory integration method based on a biomimetic sleep mechanism, which automatically organizes and consolidates the memories accumulated by the intelligent agent during the day through a timed "sleep cycle," thereby improving memory quality and retrieval efficiency.

[0005] A biomimetic sleep mechanism-based AI agent memory integration and consolidation method includes the following steps: Step S1 – Sleep Cycle Trigger: The system automatically initiates a memory integration cycle during the low-load period in the early morning each day. Step S2 – Memory Scanning and Loading: The system loads all newly added memory records for the day and sorts them by timestamp and emotional weight. Step S3 – Cross-Memory Association Discovery: Using a graph traversal algorithm and dopamine gating mechanism, semantic associations between different memory records are identified, and memories with association scores higher than a threshold are merged into memory clusters. Step S4 – Deduplication and Redundancy Elimination: Duplicate information within memory clusters is deduplicated, retaining the version with the most information. Step S5 – Abstraction Processing: Reusable pattern rules are extracted from specific event memories, and multiple similar events are abstracted into a general rule. Step S6 – Adaptive Forgetting: Memory value scores are calculated based on memory access frequency, emotional intensity, and association density; low-value memories below a threshold are either weakened or cleared.

[0006] Beneficial effects: Improved memory quality and increased retrieval accuracy after association and integration; optimized storage with deduplication and forgetting mechanisms to reduce storage redundancy; accelerated reasoning with abstract rules replacing multi-step retrieval to improve response speed; long-term learning ability with daily sleep cycles enabling the agent to continuously evolve.

Claims

1. Claim 1: A method for memory integration and consolidation in an AI intelligent agent based on a biomimetic sleep mechanism, characterized in that, Includes the following steps: The system automatically triggers a memory integration loop during periods of low system load; scans and loads all newly added memory records for the day; identifies cross-memory semantic associations using graph traversal algorithms and dopamine gating mechanisms, merging memories with associations higher than a threshold into memory clusters; deduplicates duplicate information within memory clusters, retaining the version with the most information; extracts reusable abstract rules from specific event memories; calculates value scores based on memory access frequency, emotional intensity, and association density, and memories below a threshold are either weakened or cleared.

2. Claim 2: The method according to claim 1, characterized in that, The dopamine gating mechanism mimics the regulatory effect of biological dopamine signals on memory consolidation, assigning higher consolidation priority to memories with high emotional intensity or high task success rate.

3. Claim 3: The method according to claim 1, characterized in that, The construction of the memory cluster is based on a weighted score of three dimensions: semantic vector similarity, co-occurrence frequency, and temporal proximity.

4. Claim 4: The method according to claim 1, characterized in that, The abstraction process summarizes multiple similar specific event memories into a rule-based memory such as "If condition X occurs, then perform action Y".

5. Claim 5: A biomimetic sleep mechanism-based AI intelligent agent memory integration and consolidation system, characterized in that, include: The sleep cycle scheduling module is used to trigger the integration loop at regular intervals during low-load periods; The memory scanning and loading module is used to load memory records to be processed. The association discovery engine is used to discover cross-memory associations through graph traversal and gating mechanisms; The deduplication module is used to eliminate redundant information; the abstraction engine is used to extract rules from specific events. An adaptive forgetting module is used for memory clearing based on value scores.