acceptodds
Under review as a conference paper at ICLR 2027

CORAL: Context-Oriented Retention with Anchor-Based Long-Term Memory

Abstract

Large language models (LLMs) increasingly rely on long-term conversational memory to maintain consistency across multi-turn and multi-session interactions. However, continuously accumulating uncompressed conversational histories introduces substantial redundancy, making task-relevant information increasingly difficult to retrieve and utilize. To address these challenges, we introduce CORAL (Context-Oriented Retention with Anchor-Based Long-Term Memory), a Multi-Agent System for constructing compact and retrieval-efficient conversational memory. CORAL first extracts salient information as structured anchors and then uses these anchors to guide LLM-based compression under a constrained memory budget. The resulting memory is indexed using Retrieval-Augmented Generation (RAG), enabling relevant evidence to be retrieved for downstream reasoning. By separating information preservation, compression, and query-time reasoning, CORAL aims to retain information with high future utility while removing repetitive or low-value conversational content. We evaluate CORAL on LoCoMo using question answering and event summarization, and systematically analyze factors such as compression fraction, compressor size, compression method, and model family. Overall, our results show that effective conversational memory compression requires preserving the right information, rather than simply retaining more content or using larger models. The code is anonymously available online.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Memory Compression, Large Language Models, Conversational Memory, Multi-Agent Systems, Long-Context Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_SpaaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

CORAL combines GLiNER2 anchor extraction, LLM-based compression, and retrieval-augmented QA for conversational memory. The reported results suggest that substantial compression can preserve QA performance on a small example. However, the evaluation does not establish the proposed mechanism’s benefit or support the broader empirical conclusions.

Critical strengths
  • The pipeline is simple, interpretable, and accompanied by code and detailed outputs.

  • Reporting actual compression fractions and both QA and lexical preservation metrics is useful.

Critical weaknesses
  1. The evaluation is severely insufficient and reuses the tuning data. All experiments use only the first ten questions from one LoCoMo conversation. Compression fraction and compressor size are selected using these same questions, without held-out evaluation. The headline improvement from 50% to 60% represents one additional correct answer. This undermines the empirical foundation of the method.

  2. The model-family comparison contains an implementation failure. The compression code disables thinking only for Qwen and applies the short memory budget directly to generation. In the released DeepSeek run log (experiment3_results.txt), all 100 retrieved snippet occurrences begin with unfinished compression reasoning (“Okay, so I need…”). Consequently, its 0% accuracy measures a pipeline producing truncated reasoning rather than valid compressed memories.

  3. The central methodological claim is not realized by the design. CORAL lacks explicit anchor prioritization and preservation checks. Anchors only condition the prompt, leaving the compressor to decide what survives and undermining the claimed separation between preservation and compression.

Questions for authors
  1. Do the reported gains persist on held-out conversations when compression fraction and model size are selected on a separate development set? Please provide broader evaluation with uncertainty estimates.

  2. Can you rerun the DeepSeek comparison with sufficient reasoning capacity and a separate budget for the final compressed memory? How does correcting the truncated-reasoning outputs affect the model-family conclusions?

  3. How does CORAL prioritize competing anchors under a tight budget and verify their preservation? What mechanism substantively separates preservation decisions from those already made by the compression LLM?

Reviewer_3jbLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CORAL for long-term conversational memory compression. The method first extracts salient information as structured anchors with GLiNER2. It then uses an LLM-based compressor to construct a compact memory representation under a token budget. The compressed memories are indexed and retrieved for downstream question answering. The paper evaluates compression fractions, compressor sizes, compression methods, and model families on LoCoMo. The main finding is that stronger compression or larger compressors do not necessarily improve downstream QA. CORAL also reports better QA performance than the two compression baselines under a small memory budget.

Critical strengths
  1. Separating information preservation, compression, and retrieval is a reasonable design choice. It makes the memory construction process interpretable and allows different components to be studied independently.
  2. The results show that ROUGE preservation does not necessarily correspond to downstream QA utility. This is a useful observation for evaluating compressed conversational memory.
Critical weaknesses
  1. The entire downstream QA evaluation uses only 10 questions from one LoCoMo conversation. This is especially problematic because the paper makes broad claims about long-term conversational memory. The statistical analysis also shows no significant difference for compression fraction or compressor size. The reported 60% versus 50% human accuracy therefore should not be treated as evidence that aggressive compression improves over raw memory.
  2. CORAL's main methodological claim is that extracted anchors guide compression. However, there is no direct ablation comparing CORAL with the same compressor without anchors. Without this comparison, it is unclear whether the gains come from the proposed anchor mechanism or simply from using a particular compressor and prompt. This is the most important missing experiment for validating the proposed method.
Questions for authors
  1. What is the performance of the same compression pipeline without the Preservation Agent and its anchors? This comparison is necessary to establish whether anchor-guided compression itself provides the claimed benefit.
  2. Why are only 10 questions from the first LoCoMo conversation used for the main evaluation? Can the authors report results across the full LoCoMo evaluation set and multiple conversations?
  3. How were the baselines configured under the same effective memory budget? In particular, please report the actual compression fraction for each method in Experiment 2.
  4. How should the 60% versus 50% human accuracy result be interpreted given the non-significant statistical test in Experiment 1A? The current discussion appears stronger than what the statistical evidence supports.
Reviewer_FLBiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces CORAL which is a framework for long-term conversational memory that uses extracted anchors to guide compression. The goal is to retain important information while removing redundant content. Experiments on LoCoMo show that the framework reduce memory size substantially without significantly compromising QA performance.

Critical strengths

The proposed framework is straightforward and modular, therefore, it is easy to understand and could integrate into existing memory systems.

The paper evaluates both information preservation and downstream QA performance, highlighting that retaining more information does not necessarily lead to better task performance.

Critical weaknesses

The framework largely combines existing techniques, including structured information extraction, LLM-based compression, and RAG. The technical contribution is limited, and the distinction from existing memory approaches is not sufficiently clear.

The evaluation is limited to support the paper's conclusions. Experiments use only 10 questions from a single LoCoMo conversation. Therefore, it is difficult to say there is reliable or generalizable conclusions. The paper also lacks comprehensive comparisons with existing long-term memory methods. A substantially broader evaluation is needed to establish the effectiveness of the proposed approach.

Questions for authors

The QA results for Qwen3-8B at CF=0.1 differ across Tables 1–4. Could you clarify what causes these differences?

Could you provide an ablation study? This would help clarify how much the anchor-based mechanism contributes to the reported performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.