Skip to content

Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction

Conference: AAAI 2026 Oral
arXiv: 2511.13410
Code: GitHub
Area: Information Retrieval Keywords: Personalized Dialogue, Long-term Memory, Hierarchical Memory, User Modeling, Dialogue Assistant

TL;DR

This paper proposes H2Memory, a four-layer hierarchical heterogeneous memory structure (Log Graphs / Background Memory / Topic Outlines / Principles), validated on the PAL-Set dataset (100 users × 8.4 months of interaction), improving BLEU-1 on demand paraphrasing and solution recommendation tasks from 13.59 to 26.67.

Background & Motivation

Background: Personalized dialogue systems must retain users' long-term preferences and history. Existing methods (RecurSum, MemoryBank, etc.) primarily rely on simple summarization-based memory.

Limitations of Prior Work: (a) User preferences emerge from fragmented logs rather than explicit statements; (b) a single memory structure cannot simultaneously handle information at different granularities; (c) long-term personalized dialogue datasets are lacking.

Key Challenge: A semantic gap exists between users' daily behavioral logs (fragmented, implicit) and dialogue-level demands (structured, explicit).

Goal: Construct multi-granularity memory from users' long-term logs and conversation history to support personalized assistants in demand understanding and solution recommendation.

Key Insight: Four memory layers progressively aggregate information from raw logs to abstract principles, retrieving user information at different granularities from each layer.

Core Idea: A hierarchical heterogeneous memory structure (Log Graphs → Background → Topics → Principles) progressively abstracts fragmented user behavior into retrievable personalized knowledge.

Method

Overall Architecture

Input: user log stream (behavioral records) + multi-turn conversation history. Output: personalized demand understanding / solution recommendation. H2Memory consists of four layers: (1) Log Graphs → (2) Background Memory → (3) Topic Outlines → (4) Principles.

Key Designs

  1. Log Graphs:

    • Function: Construct situational descriptions from fragmented logs.
    • Mechanism: Related log fragments are connected via relational edges (two types: Caused_by and Follows) to form structured situational graphs; a situation description is then generated for each connected subgraph.
    • Design Motivation: Raw logs are too fragmented for direct retrieval and must be aggregated into meaningful situational units.

  2. Background Memory:

    • Function: Maintain long-term summaries of fixed aspects (e.g., occupation, interests) across time.
    • Mechanism: Recursively updated — summaries of relevant aspects are revised upon each new log entry.
    • Design Motivation: Provides a persistent representation of the user profile.

  3. Topic Outlines:

    • Function: Extract structured information on demands, solutions, and preferences from conversations.
    • Mechanism: Each dialogue turn yields triples (demand, solution, preference), organized by topic.
    • Design Motivation: Demands and preferences expressed in conversation are central to personalization.

  4. Principles:

    • Function: Abstract high-level demand types from topic outlines.
    • Mechanism: KMeans clustering is applied to categorize demands and generalize them into universal principles.
    • Design Motivation: Provides the highest-level personalized prior.

Loss & Training

During retrieval, the top \(k=3\) most similar entries (via paraphrase-multilingual-mpnet embeddings) are fetched from each of the four layers; the aggregated context is then fed into an LLM for response generation.

Key Experimental Results

Main Results

Method Demand Paraphrase BLEU-1 Solution Selection Score GPT-4 Score
Vanilla (no logs) 13.59 18.95 17.50
RecurSum — — Lower
MemoryBank — — Lower
H2Memory 26.67 38.32 32.54

Ablation Study

Configuration Impact Note
w/o Topic Memory (\(M_T\)) ~20% drop Most critical for demand understanding
w/o Background Memory (\(M_B\)) ~6% drop Important for preference alignment
w/o Principles Minor drop Provides high-level prior

Key Findings

  • H2Memory improves demand paraphrase BLEU-1 from 13.59 to 26.67 (a 96% relative gain).
  • Topic Memory contributes most to demand understanding (~20% drop when removed); Background Memory contributes to preference alignment (~6% drop).
  • PAL-Set achieves high quality in human evaluation: logs 2.75/3.0, dialogues 2.67/3.0.
  • Different memory layers serve distinct tasks, validating the necessity of heterogeneous design.

Highlights & Insights

  • Systematic four-layer heterogeneous memory design: Progressive aggregation from raw logs to abstract principles mirrors the "episodic–semantic–procedural" hierarchy of human memory.
  • PAL-Set dataset: 100 users × 29 sessions × 996 logs × 8.4-month span, filling a gap in long-term personalized dialogue benchmarks.
  • Ablation validates design rationale: The independent contribution of each memory layer is clearly quantifiable.

Limitations & Future Work

  • Retrieval strategy is relatively simple: A fixed \(k=3\) per layer is used, with no dynamic or adaptive retrieval.
  • Dataset is LLM-synthesized: Although human-validated, it may not fully reflect real user behavior.
  • Memory forgetting/decay is not considered: In long-term interactions, outdated information may need to be discarded.
  • Future work could introduce proactive memory update strategies (e.g., handling conflicts between new logs and existing memory).
  • Cross-dataset validation on EMG-RAG demonstrates that H2Memory generalizes beyond PAL-Set, confirming the broader applicability of the design.
  • vs. RecurSum: RecurSum employs recursive summarization, whereas this work uses a four-layer heterogeneous structure providing richer memory representations.
  • vs. MemoryBank: MemoryBank relies on a single memory structure; the hierarchical design here offers greater flexibility.
  • vs. Persona-based Dialogue: Traditional persona methods use fixed descriptive sentences, whereas H2Memory employs dynamically updated multi-granularity memory, better suited for evolving user preferences in long-term interaction.
  • Insight: Personalization systems require user knowledge at varying granularities; a single representation is insufficient.
  • Insight: Automated management of heterogeneous memory (when to merge, when to forget) remains a key open problem in personalized dialogue.

Rating

  • Novelty: ⭐⭐⭐⭐ Systematic and well-motivated four-layer heterogeneous memory design
  • Experimental Thoroughness: ⭐⭐⭐⭐ Includes ablation studies and human evaluation
  • Writing Quality: ⭐⭐⭐⭐ Clear and well-structured
  • Value: ⭐⭐⭐⭐ Practical value for long-term personalized dialogue

Additional Notes

  • The methodology and experimental design of this work offer meaningful reference for related research areas.
  • Future work may validate generalizability and scalability across more scenarios and at larger scale.
  • Potential research value exists in combining this work with recent related approaches (e.g., intersections with RL/MCTS or multimodal methods).