acceptodds
Under review as a conference paper at ICLR 2027

Isolated Facts Do Not Tell the Whole Story of a Long-Term Conversation

10 views

Abstract

Long-horizon conversations do not preserve a user state as independent facts: evidence is distributed across sessions, revised over time, and often relevant only through intermediate memories. Existing memory systems recover these dependencies via costly mechanisms: graph-structured memories construct elaborate relational structures during indexing and rely on multi-stage graph traversal at query time, while agentic retrievers repeatedly invoke LLMs to plan and refine retrieval. We propose TraceMem, a lightweight alternative that organizes conversational histories into an evolving flat memory. At retrieval time, TraceMem uses two complementary relations: the intra-relation links historical and current expressions of the same memory to preserve state evolution, while the inter-relation connects memories through bridge entities to recover cross-session dependencies. Together they form complete memory trajectories, from which multi-hop and temporal questions are answered without complex graph reasoning or LLM-guided iterative retrieval. Across LoCoMo-10 and LongMemEval-S, TraceMem outperforms agentic baselines by 6.1% while achieving near-zero-latency retrieval, demonstrating that structured evidence reconstruction does not require a complex query-time reasoning system.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Memory Management, RAG, Agentic AI, Conversational Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vGPzOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TraceMem, a conversational memory system that stores propositions and preserves their revision histories. Retrieval combines intra-relations between historical and current versions with inter-relations based on shared entities, then returns the supporting dialogue turns to a reader. Experiments on LoCoMo-10 and LongMemEval-S report improved answer accuracy and evidence recall, together with low retrieval latency. The paper also includes relation ablations and hyperparameter studies.

Critical strengths
  • A practical and intuitive memory retrieval design. The combination of intra- and inter-relations provides a relatively simple way to recover distributed evidence while preserving memory evolution. Avoiding iterative LLM calls at retrieval time is particularly attractive for latency-sensitive applications.

  • Promising empirical results with useful supporting analysis. TraceMem achieves strong results on two established benchmarks. The component ablations and evidence-recall experiments further support the usefulness of the proposed relations, rather than relying solely on final QA accuracy.

Critical weaknesses
  • Insufficient isolation of the core retrieval contribution. Although the component ablations demonstrate the benefits of intra- and inter-relations, the paper lacks a comparison with simpler entity-based expansion over the same proposition store. It remains unclear how much of the improvement comes specifically from trajectory scoring rather than memory construction and relation-based expansion.

  • Efficiency and dynamic-memory claims need stronger validation. The reported retrieval latency is promising, but its measurement scope and scaling behavior remain unclear. Although the benchmarks include temporal and knowledge-update questions, the paper lacks direct evaluation of progressive memory updates and their computational costs.

Questions for authors
  1. How does TraceMem compare with simpler entity-based expansion over the same proposition store, using identical embeddings and reader context budgets?

  2. Does the reported 0.08-second latency include all retrieval operations, including query embedding, trajectory scoring, and candidate embedding? How does retrieval latency scale with memory size? Also, Figure 3 appears to show a smaller offline token footprint for LightMem than TraceMem. Could the authors clarify this discrepancy?

  3. Could the authors provide a more direct evaluation of incremental memory updates, particularly when previously stored facts are revised or contradicted?

  4. How sensitive is TraceMem to the memory construction model? Since proposition extraction and consolidation rely on Llama-3.3-70B, would using a smaller model significantly affect retrieval accuracy and indexing efficiency?

Reviewer_Mjb2OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TraceMem, a lightweight framework for long-term conversational memory. It incrementally organizes dialogue into semantic and episodic propositions while preserving historical memory states. At retrieval time, intra-relations connect historical and current memories, while inter-relations recover cross-session dependencies through shared entities. The resulting memory trajectories are mapped back to source dialogue turns for answer generation. Experiments on LoCoMo-10 and LongMemEval-S report improved accuracy over existing memory systems with low online retrieval latency.

Critical strengths

1.Effective memory retrieval design. The combination of intra- and inter-relations provides a simple way to handle evolving user states and cross-session dependencies. The ablation results in Section 4.3 support the complementary roles of these two relations.

2.Promising accuracy–efficiency trade-off. TraceMem achieves competitive results on both LoCoMo-10 and LongMemEval-S while substantially reducing online retrieval latency compared with agentic baselines.

Critical weaknesses

1.Limited novelty and unclear contribution of trajectory retrieval. The method combines established techniques such as proposition consolidation, entity-based linking, and beam search. The current ablations do not fully isolate the benefits of trajectory-level retrieval from those of the underlying memory representation.

2.Incomplete evidence for efficiency and scalability. While online retrieval is fast, the paper provides limited analysis of indexing and retrieval costs as memory grows. Moreover, the claim of the smallest offline token footprint in Section 4.2 is inconsistent with Figure 3.

Questions for authors

1.Trajectory retrieval (Section 3.3): How much of the improvement comes from trajectory construction itself? Could the authors compare against simpler hybrid retrieval or entity-based expansion using the same consolidated memory?

2.Evaluation metrics (Appendix D, Tables 5–6): Why do F1 and BLEU-1 exhibit nearly constant differences across many methods and categories (e.g., exactly 3.00 points across all four LongMemEval-S categories for TraceMem)? Could the authors clarify the metric implementations?

3.Efficiency and scalability (Section 4.2, Figure 3; Appendix A): Why is TraceMem described as having the smallest offline footprint when LightMem reports a lower cost? How do retrieval latency and memory-update costs scale with increasing conversation history?

Reviewer_EFrcOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces TraceMem, a lightweight framework for long-term memory that structures conversations as evolving propositions. It employs intra-relations to maintain historical memory states and inter-relations to link evidence across sessions. Experimental results on LoCoMo-10 and LongMemEval-S show that TraceMem achieves better retrieval accuracy and efficiency than existing memory baselines.

Critical strengths

Efficient retrieval design. TraceMem avoids iterative LLM calls during retrieval, significantly reducing online latency and context consumption.

Useful component analysis. The ablations demonstrate the complementary contributions of intra- and inter-relations.

Critical weaknesses
  • In Tables 5 and 6, F1 scores are consistently and [accurately 3.0 points higher] than BLEU-1 across different methods and categories. This unusual pattern needs further explanation and raises concerns about metric reliability. If solved, I'm very glad to discuss more and raise the score.

  • Historical memory consolidation and entity-based evidence linking have been explored in prior work. The proposed framework mainly combines existing techniques, offering limited methodological novelty.

  • Scoring all stored propositions and expanding trajectories through shared entities may limit scalability. However, the paper lacks an analysis of retrieval latency and memory consumption as conversation histories grow.

Questions for authors
  • Can the authors explain the nearly constant F1–BLEU-1 differences in Tables 5 and 6 and provide the evaluation scripts?

  • How does TraceMem's retrieval latency scale with increasing memory size?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.