An agent long-term memory system based on atomic facts and hierarchical structure

By using an agent long-term memory system based on atomic facts and hierarchical structure, the problems of information loss and redundant storage in existing technologies are solved, achieving efficient and stable long-term memory management and improving the performance of large language model agents in long-term interactions.

CN122222008APending Publication Date: 2026-06-16UNIV OF SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
UNIV OF SCI & TECH OF CHINA
Filing Date
2026-03-16
Publication Date
2026-06-16

AI Technical Summary

Technical Problem

Existing memory enhancement technologies suffer from problems such as loss of fine-grained information, redundant storage of dialogue content, and uncontrolled dynamic updates, leading to a decline in the reliability and efficiency of long-term memory systems.

Method used

A long-term memory system for intelligent agents based on atomic facts and hierarchical structure is adopted. Through denoising processing by a large language model extractor with fine-tuned instructions and lightweight semantic reasoning, high-information-density atomic facts are generated. Event blocks are constructed through similarity retrieval and dynamic aggregation. Combined with a session-level batch processing and hierarchical hybrid retrieval mechanism, stable memory management is achieved.

Benefits of technology

It increases information density, reduces redundant storage and noise interference, maintains the contextual integrity of dialogue, improves retrieval efficiency and reasoning accuracy, provides personalized long-term memory management capabilities, and enhances the reliability and stability of large language model agents in long-term interactions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122222008A_ABST
    Figure CN122222008A_ABST
Patent Text Reader

Abstract

The application provides an agent long-term memory system based on atomic facts and a hierarchical structure, and the system comprises: a memory storage module, which is used for denoising and lightweight semantic reasoning on an original dialogue stream through a large language model fine-tuned by instructions as an atomic fact extractor, extracting atomic facts with independent semantics and capable of being understood individually, and structurally packaging the atomic facts as minimum semantic units for storage; a memory updating module, which is used for obtaining candidate historical facts related to a current fact through a similarity retrieval mechanism; and a memory retrieval module, which is used for generating a response by fusing a large language model after input according to current query content. The system extracts high information density atomized facts from an original dialogue stream, constructs cross-temporal long-range semantic associations, thereby greatly reducing storage redundancy, and realizes efficient and stable long-term memory storage and retrieval.
Need to check novelty before this filing date? Find Prior Art