PaperlayerView exact revision on arXiv

Source-linked AI summary

CueMem: Cue-Guided Context Reconstruction for Long-Term Conversational Memory

Changjian Wang, Rongzhen Li, Weili Guan, Shuming Shi, Quan Lu, Ning Jiang

arXiv:2609.12354v1cs.CL

TL;DR

Long-term conversational agents need to answer queries from extended dialogue histories, but full histories are costly and unreliable, and compressed memories may lose fine-grained evidence. CueMem retrieves source-linked cues and expands from their turn anchors to reconstruct relevant dialogue context. It consistently outperforms representative memory baselines on LoCoMo and LongMemEval while using fewer query-time tokens and less latency than the full-history setting.

  • Problem

    Full dialogue histories can be costly and unreliable, while compressed memory units may lose fine-grained evidence needed to answer queries.

    Source: §1 · p. 1

  • Method

    CueMem treats fine-grained, source-linked memories as retrieval cues and expands from their source-turn anchors over a turn graph to reconstruct query-relevant dialogue context.

    Source: §1 · p. 2, §4.1 · p. 3, §4.2 · p. 4

  • Results

    CueMem consistently outperforms representative long-term memory baselines on LoCoMo and LongMemEval while reducing query-time token consumption and latency compared with the full-history setting.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Cue-guided context reconstruction offers a simple and scalable strategy for long-term conversational memory.

    Source: §6 · p. 8

    CueMem primarily retrieves and reconstructs dialogue evidence, while some multi-hop questions require integrating that evidence with external knowledge.

    Source: §5.2 · p. 7

Abstract

from arXiv · show

Long-term conversational agents must answer user queries by recalling information from extended dialogue histories, yet directly using the full history is costly and often unreliable, while compressed memory units may lose fine-grained evidence needed for question answering. Motivated by the reconstructive view of autobiographical memory, we propose CueMem, a cue-guided framework that treats extracted memory records as retrieval cues rather than self-contained evidence and reconstructs query-relevant dialogue context from their source turns. During memory construction, CueMem extracts fine-grained memory cues from dialogue turns and links each cue to its source turn. At query time, it retrieves query-relevant cues, maps them to source-turn anchors, and expands from these anchors over a turn graph that captures temporal proximity and semantic relatedness, reconstructing a compact evidence context from the original dialogue for LLM answer generation. Experiments on LoCoMo and LongMemEval show that CueMem consistently outperforms representative long-term memory baselines. Further analyses show that graph-based context reconstruction helps recover supporting dialogue evidence while reducing query-time input tokens and latency compared with the full-history LLM setting. These results highlight retrieval cues as an effective alternative to self-contained memory evidence for long-term conversational question answering.

1 Introduction

Full dialogue histories can be costly and unreliable, while compressed memory units may obscure fine-grained evidence. CueMem instead uses query-matched cues to recover relevant dialogue context from their source turns.

1 Introduction

  • Full interaction histories can be ineffective because of context-length limits, noisy information, lost-in-the-middle effects, and increased inference latency.Source: §1 · p. 1
  • Compressed or decontextualized memory units may entangle multiple semantic signals, making fine-grained query evidence difficult to retrieve.A single “travel planning” unit, for example, may combine destination, budget, and hotel preferences.Source: §1 · pp. 1–2
  • CueMem extracts fine-grained cues from dialogue turns, links each cue to its source turn, and expands from those anchors over a turn graph to gather query-relevant context.The graph connects turns by temporal proximity and semantic similarity; the reconstructed context is supplied to the LLM for answer generation.Source: §1 · p. 2
  • Using the reconstructive view of autobiographical memory, CueMem treats retrieved memories as cues for reconstructing dialogue context, not as self-contained evidence.The proposed process first identifies query-matched fine-grained cues, then recovers the surrounding original dialogue before answer generation.Source: §1 · p. 2
  • CueMem consistently outperforms strong long-term memory baselines on two conversational question-answering benchmarks.The benchmarks are LoCoMo and LongMemEval.Source: §1 · p. 2

2 Related Work

Related work retrieves dialogue or memory records to support long-context and conversational reasoning, but fixed chunks and compressed records can mismatch the fine-grained needs of dynamic conversations. Autobiographical-memory research motivates treating cues as guides for reconstructing evidence.

2 Related Work

3 Problem Formulation

The task is to answer a query faithfully from a long dialogue history by selecting a compact, relevant evidence context. The answer-generation prompt conditions the LLM on the query and that context.

3 Problem Formulation

  • The task is to generate an answer relevant to query q and faithful to dialogue history D by selecting a query-relevant evidence context Eq ⊆ D.The history is represented as a chronological sequence of turns and may span multiple sessions; each turn contains a user utterance and the corresponding agent response.Source: §3 · p. 3
  • Because dialogue history can be long, noisy, and only partly relevant, the system should select compact evidence rather than input the entire history.Source: §3 · p. 3
  • The answer is generated by conditioning on query q and evidence context Eq in an answer-generation prompt for the LLM.Eq contains the historical information needed to answer q.Source: §3 · p. 3

4 Methodology

CueMem represents dialogue as fine-grained cues linked to source turns, then uses query-relevant cues to anchor reconstruction of evidence context from a turn graph.

4.1 Memory Cue Extraction

  • Each memory cue is an LLM-extracted relational triple linked to its source turn and encoded as a dense vector for retrieval and graph construction.A dialogue turn may yield multiple cues or none.Source: §4.1 · p. 3

4.2 Turn Graph Construction

  • The turn graph combines fixed-window temporal neighbors with non-local semantic links formed by matching cue-level nearest neighbors and mapping them to source turns.Cue-level similarity uses atomic representations to reduce interference from mixed semantics.Source: §4.2 · p. 4

4.3 Context Reconstruction

  • At query time, retrieved cues supply anchors; one-hop temporal and semantic neighbors join them as evidence turns, which are ordered chronologically before generation.The selected graph neighborhood includes local dialogue context and semantically related non-local turns while limiting irrelevant history.Source: §4.3 · pp. 4–5

4.4 Memory Management

  • CueMem separates offline cue extraction and graph construction from online retrieval and reconstruction; additions are incremental, deletion uses forgetting policies, and updates add newer cues.The method introduces no specialized update mechanism and instructs generation to prioritize more recent evidence.Source: §4.4 · p. 5

5 Experiments

Across LoCoMo and LongMemEval, CueMem achieves the strongest overall accuracy among the compared systems. Ablations show accuracy drops when graph expansion or the priority instruction is removed, while comparison with full-history prompting shows lower token use and latency alongside better accuracy.

5.1 Experimental Settings

  • Evaluation uses LLM-judged accuracy on LoCoMo’s 10 conversations and 1,540 questions and LongMemEval-S’s 500 questions, with Llama-3.3-70B-Instruct shared across methods.Source: §5.1 · pp. 5–6

5.2 Main Results

5.3 Ablation Study

  • On LoCoMo, disabling graph expansion lowers overall accuracy from 81.1% to 71.4%, showing that retrieved anchor turns alone often lack sufficient supporting evidence.Source: Figure 3 · p. 7, §5.3 · p. 7
  • Removing the priority instruction lowers overall accuracy from 75.2% to 72.4%, knowledge-update accuracy from 87.2% to 80.8%, and temporal-question accuracy by 6.8%.Source: Figure 3 · p. 7, §5.3 · p. 8

5.4 Efficiency Analysis

  • Against full-history prompting, CueMem uses about 10% of its input tokens on LoCoMo and less than 2% on LongMemEval, while reducing latency by 42.5% and 87.4%, respectively, and improving accuracy.Source: Table 3 · p. 8, §5.4 · p. 8

6 Conclusion

CueMem treats extracted memories as fine-grained retrieval cues linked to source turns, then reconstructs query-relevant dialogue context over a turn graph. On LoCoMo and LongMemEval, it outperforms representative long-term memory baselines while reducing query-time tokens and latency compared with the full-history setting.

  • CueMem consistently outperforms representative long-term memory baselines on LoCoMo and LongMemEval while reducing query-time token consumption and latency compared with the full-history setting.Source: §6 · p. 8
  • CueMem treats extracted memories as fine-grained retrieval cues, links them to source turns, and expands over a turn graph to reconstruct query-relevant dialogue context.Source: §6 · p. 8
  • The authors present cue-guided context reconstruction as a simple and scalable strategy for long-term memory in conversational agents.Source: §6 · p. 8
Loading 2609.12354v1…