Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HiMeS: Hippocampus-inspired Memory System for Personalized AI Assistants

Hailong Li, Feifei Li, Wenhui Que, Xingyu Fan

TL;DR

HiMeS addresses memory limitations in retrieval-augmented generation for personalized AI assistants in industrial settings. It fuses a short-term memory extractor trained with reinforcement learning to compress dialogue and proactively retrieve documents with a partitioned long-term memory that stores user histories and re-ranks content via an attention-inspired mechanism. Across a real-world industrial dataset, HiMeS improves QA and contextual alignment, with ablations showing both memory modules are essential and the HSER RL reward yields strong performance. The framework demonstrates adaptability across different LLM backends, offering a practical, plug-and-play memory layer for reliable, personalized knowledge services.

Abstract

Large language models (LLMs) power many interactive systems such as chatbots, customer-service agents, and personal assistants. In knowledge-intensive scenarios requiring user-specific personalization, conventional retrieval-augmented generation (RAG) pipelines exhibit limited memory capacity and insufficient coordination between retrieval mechanisms and user-specific conversational history, leading to redundant clarification, irrelevant documents, and degraded user experience. Inspired by the hippocampus-neocortex memory mechanism, we propose HiMeS, an AI-assistant architecture that fuses short-term and long-term memory. Our contributions are fourfold: (1) A short-term memory extractor is trained end-to-end with reinforcement learning to compress recent dialogue and proactively pre-retrieve documents from the knowledge base, emulating the cooperative interaction between the hippocampus and prefrontal cortex. (2) A partitioned long-term memory network stores user-specific information and re-ranks retrieved documents, simulating distributed cortical storage and memory reactivation. (3) On a real-world industrial dataset, HiMeS significantly outperforms a cascaded RAG baseline on question-answering quality. (4) Ablation studies confirm the necessity of both memory modules and suggest a practical path toward more reliable, context-aware, user-customized LLM-based assistants.

HiMeS: Hippocampus-inspired Memory System for Personalized AI Assistants

TL;DR

HiMeS addresses memory limitations in retrieval-augmented generation for personalized AI assistants in industrial settings. It fuses a short-term memory extractor trained with reinforcement learning to compress dialogue and proactively retrieve documents with a partitioned long-term memory that stores user histories and re-ranks content via an attention-inspired mechanism. Across a real-world industrial dataset, HiMeS improves QA and contextual alignment, with ablations showing both memory modules are essential and the HSER RL reward yields strong performance. The framework demonstrates adaptability across different LLM backends, offering a practical, plug-and-play memory layer for reliable, personalized knowledge services.

Abstract

Large language models (LLMs) power many interactive systems such as chatbots, customer-service agents, and personal assistants. In knowledge-intensive scenarios requiring user-specific personalization, conventional retrieval-augmented generation (RAG) pipelines exhibit limited memory capacity and insufficient coordination between retrieval mechanisms and user-specific conversational history, leading to redundant clarification, irrelevant documents, and degraded user experience. Inspired by the hippocampus-neocortex memory mechanism, we propose HiMeS, an AI-assistant architecture that fuses short-term and long-term memory. Our contributions are fourfold: (1) A short-term memory extractor is trained end-to-end with reinforcement learning to compress recent dialogue and proactively pre-retrieve documents from the knowledge base, emulating the cooperative interaction between the hippocampus and prefrontal cortex. (2) A partitioned long-term memory network stores user-specific information and re-ranks retrieved documents, simulating distributed cortical storage and memory reactivation. (3) On a real-world industrial dataset, HiMeS significantly outperforms a cascaded RAG baseline on question-answering quality. (4) Ablation studies confirm the necessity of both memory modules and suggest a practical path toward more reliable, context-aware, user-customized LLM-based assistants.
Paper Structure (23 sections, 1 equation, 9 figures, 6 tables, 2 algorithms)

This paper contains 23 sections, 1 equation, 9 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Short-term memory and dialogue compression
  4. Long-term memory and memory storage
  5. Method
  6. Short-term memory module
  7. Long-term memory module
  8. Partitioned memory storage
  9. Attention-inspired Rerank
  10. Experiments
  11. Dataset
  12. Metrics
  13. Baseline
  14. Modular ablation experiments
  15. Post-training ablation experiments
  16. ...and 8 more sections

Figures (9)

  • Figure 1: Illustration of HiMeS
  • Figure 2: RLHF Training Pipeline of Query Rewriter
  • Figure 3: Storage, pre-recall, and reordering mechanisms of Attention-inspired-rerank
  • Figure 4: Case of HiMeS vs. Native RAG
  • Figure 5: Results under different $\lambda$ values
  • ...and 4 more figures