acceptodds
Under review as a conference paper at ICLR 2027

EvoMem: Type-Adaptive Retrieval over Versioned Memory for Self-Evolving LLM Agents

Abstract

Long-term memory enables LLM agents to retain user-specific information across extended interactions. However, maintaining evolving user attributes requires identifying their current values and their historical states at specified times. Existing approaches either retrieve outdated values from an insufficiently updated knowledge base or compress observations into summaries that obscure revision histories. Moreover, most systems tightly couple memory storage and retrieval, thereby imposing a fixed access path regardless of query requirements. To address these limitations, we introduce EvoMem, a memory architecture that supports both current-state and historical reasoning by decoupling versioned storage from query-adaptive retrieval. EvoMem preserves complete histories of user attributes without committing them to a single retrieval structure, and dynamically selects and ranks evidence according to the query type. This design poses two key challenges. Different queries require distinct retrieval mechanisms, while retaining all attribute versions may cause obsolete information to compete with current values and lead to incorrect retrieval. EvoMem addresses these challenges through type-adaptive retrieval and self-evolving memory consolidation. Type-adaptive retrieval selects appropriate memory structures and retrieval paths for factual, time-sensitive, and multi-session reasoning queries. Self-evolving memory consolidation integrates Versioned Knowledge Evolution (VKE), which preserves timestamped attribute versions, with Error-Driven Pattern Learning (EDPL), which extracts reusable patterns from prior retrieval errors to continually refine ranking policies. Experiments on LongMemEval and LoCoMo demonstrate that EvoMem consistently outperforms strong baselines under predicted query-type routing. In particular, on LoCoMo, EvoMem exceeds the strongest baseline by points in QA accuracy and points in token-level F1. The source code is available at https://anonymous.4open.science/r/Evomem/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Term Memory, LLM Agents, Retrieval-Augmented Generation, Temporal 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_LEDxOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EvoMem stores dated versions of facts and uses query types and development feedback to choose retrieval rules for questions about changing information.

Critical strengths

The dated version chains keep old facts available for questions about earlier states.

Critical weaknesses
  1. It specifies 400 test questions, and Appendix B.4 describes averaged binary scores. Accuracy must therefore change in steps of 0.25. Table 1 reports 82.6 and 88.4, neither obtainable after rounding those outcomes. Another aggregation rule might explain this, but the description leave it unresolved. Provide actual denominators and per-question verdicts before interpreting the reported gains.

  2. It can accept a count of 1 instead of 2, or a week instead of an exact date. The pre-check also accepts high keyword overlap. These rules can reward errors in abilities tested by LongMemEval, compare verdicts with human judgments. The current rules makes memory accuracy difficult to trust; rescore using question-specific factual requirements.

  3. It train EDPL on known development answers, then freeze it before testing. This supports offline tuning, but not continued learning from noisy user feedback. EvoMem needs a comparable feedback–update–evaluation sequence. The development curve alone do not establish continued improvement; report later gains and new errors.

  4. It places a wish to adopt a child and an interest in counseling within one career history. These may be separate attributes, rather than changing versions of one fact. This example raises a core extraction concern, without establishing its frequency. checks claims against sources. EvoMem’s version chains needs similar checks of attribute identity and update relations.

  5. It promises the same first 30,000 characters, but Figure 4 gives EvoMem earlier facts outside FullText’s window. Clarify whether this example follows the scored protocol. Extra information could explain its advantage.The comparison need exact input records to separate versioning benefits from differences in evidence access.

  6. These results demonstrate that, In contrast, a broken sentence.

Questions for authors

Please refer to the weaknesses.

Reviewer_Vx26OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies long-term memory in LLM agents. The authors argue that existing approaches either retrieve outdated information or compress observations into summaries that lose historical information, and typically use fixed retrieval strategies regardless of the query. To address these limitations, they propose EvoMem, a memory system that preserves historical versions of user attributes and selects retrieval strategies based on the query type. The method adjusts its retrieval policy based on previous errors. Experimentally, the authors report improvements over prior methods on LoCoMo and LongMemEval.

Critical strengths
  1. The problem of long-term memory is relevant.
  2. The method performs favorably over the reported baselines.
Critical weaknesses

My main criticism of work is the added complexity for the relatively small gains the method yields. Concretely:

  • The novelty over prior work is unclear. Although the authors cite MemoTime and MemGuide, it's unclear what the proposed combination adds beyond existing approaches to temporal memory and query-adaptive retrieval. At the same time, the comparison with prior work is incomplete, e.g. EverMemOS is missing, despite being a strong recent method.
  • The improvements over the strongest baselines are relatively small (eg +2.0 on LongMemEval), considering the technical complexity of the method.
  • The evaluation protocol is concerning: the LLM judge explicitly allows numerically incorrect answers (e.g., 1 instead of 2). Results are also reported for a single seed without confidence intervals.
  • The proposed self-evolving component is trained using feedback on a development set and frozen during evaluation. Thus, the experiments do not directly demonstrate continual improvement during deployment.
  • Evaluation uses only 400 questions from LoCoMo, despite the benchmark containing 1,540 non-adversarial questions. It's not clear why this choice was made.

Additionally, the results potentially aren't statistically significant, as the authors themselves write: "Results are reported for a single backbone (DeepSeek-V4-Flash), a single seed (42), and a single judge model (GLM-5.3-Flash), without confidence intervals or significance tests.".

Questions for authors

See above.

Reviewer_7ZsvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EvoMem, an agent memory framework that decouples versioned storage from query-adaptive retrieval to track changing user attributes across multi-session interactions. The architecture builds four complementary memory structures including a four-layer Fact Graph that maintains chronological version chains of attributes without overwriting historical records. At query time, a lightweight router infers query requirements across factual, temporal, and inferential dimensions to dynamically activate specific retrieval paths and score candidates using heuristic beam search. The system also includes an error-driven pattern learning mechanism that uses feedback from development errors to adjust retrieval policies offline. The authors evaluate EvoMem on 500-question subsets of LoCoMo and LongMemEval using DeepSeek-V4-Flash, reporting gains in accuracy and token-level F1 over several memory baselines.

Critical strengths
  1. Decoupling persistent versioned storage from query-adaptive retrieval is a well-motivated concept. Maintaining explicit version chains in the Fact Graph avoids overwriting historical records while allowing query-specific routing to filter out obsolete values.

  2. The modular design of the ingestion structures and retrieval profiling is clear and systematically formulated. Combining explicit facts, full session embeddings, and versioned graphs provides a solid foundation for handling diverse query types.

Critical weaknesses

Even though i really like the approach proposed by this paper, there are several weaknesses in experimental validation:

  1. While the authors acknowledge this limitation in Appendix D, relying entirely on a single random seed and a single backbone model without reporting standard deviations makes it impossible to verify the statistical significance of the results given the high variance of LLM outputs.

  2. The evaluation scope is heavily restricted to small 500-question subsets of conversational question-answering benchmarks. It completely omits dynamic interactive agent environments with tool calls or multi-turn actions, failing to validate whether the memory system works in realistic agent deployments.

  3. The paper does not provide an adequate comparison of memory construction costs and execution latency against competing baselines. Ingestion requires a full-session LLM call for every session via Intent and Preference Inference, resulting in high token costs that are only compared against LightMem while omitting other key baselines and all wall-clock latency numbers.

  4. Crucial hyperparameters and scoring weights across the beam search and pruning modules are fixed without any sensitivity analysis. Without empirical validation across varying threshold ranges, it remains unclear whether these hand-crafted settings overfit the small evaluation split.

Questions for authors
  1. Could you provide results across at least three distinct random seeds and multiple backbone models with standard deviations to confirm statistical significance?

  2. How does EvoMem perform on interactive agent benchmarks that require tool use and sequential action execution rather than purely conversational question answering?

  3. What is the average wall-clock latency of the ingestion and retrieval pipelines, and how does the total token overhead for memory construction compare against other baselines?

  4. How sensitive is retrieval performance to changes in the pruning thresholds, decay rates, and smoothing counts defined in Section 3.2?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.