acceptodds
Under review as a conference paper at ICLR 2027

Melo: A Cognitive Science–Inspired Closed-Loop Long-Term Memory Framework for AI Agents

Abstract

Long-term memory enables AI agents to maintain coherent and personalized behavior across sessions. However, many memory architectures implement writing, retrieval, revision, and forgetting as separate operations, leaving weak attribution between a memory's use and later evidence that confirms or contradicts it. Drawing on cognitive science, we propose Melo, a closed-loop framework that coordinates attention-biased selective encoding, outcome-triggered feedback admission, evidence-linked procedural reconsolidation, and time-dependent retention control. Corrections, tool evidence, task outcomes, and conflict signals can therefore change the memory state used for subsequent retrieval and response generation. We also introduce CMLM-Assist, a 1,536-question benchmark that evaluates user-profile retention, cross-session task continuity, procedural reuse, and memory-grounded response generation. Across six baselines and three language-model backbones, Melo achieves competitive overall question-answering performance on the public LoCoMo benchmark and state-of-the-art performance under the evaluated CMLM-Assist protocol. These results support integrated coordination of memory use and revision while delimiting unresolved challenges in conflict resolution, equal-budget comparison, and long-term deployment.

Primary area:
foundation or frontier models, including LLMs
Keywords:
AI agents, closed-loop memory, cognitive science, long-term memory, memory consolidation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7tHJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Melo, a closed-loop long-term memory framework for LLM agents motivated by cognitive accounts of attentional bias, prediction error, and reconsolidation. The core idea is that evidence observed after a memory-supported response (user corrections, tool outputs, task outcomes) should be returned to the persistent state used by later queries. The implementation is a deterministic, rule-based pipeline: a write score computed from adapter-supplied importance, confidence, profile overlap, and noise tags; tag-based routing into working, episodic, semantic, and procedural memory; a binary source/tag trigger for feedback admission; consolidation of three or more procedure events into one workflow summary; and exponential time-decay retention. The authors also construct CMLM-Assist, a synthetic, template-based benchmark with 1,920 questions (1,536 held out). Experiments on LoCoMo and CMLM-Assist with DeepSeek-V4-Flash, Qwen3-14B, and GPT-4o-mini show that Melo achieves the highest F1, BLEU-1, ROUGE-L, and METEOR against six baselines (Vanilla RAG, Mem0, A-Mem, MemoryOS, MemoryBank, MemGPT/Letta), and a matched GPT-4o-mini ablation shows that removing each of four components lowers F1.

Critical strengths

Clear problem definition and transparent scoping. The paper gives a precise four-condition definition of a closed loop (memory use -> feedback admission -> state change -> later consultation) and separates it from outcome metrics such as accuracy and retention. The Monday -> Friday deadline example makes the target failure mode concrete. The authors are also candid about the limits of their evidence, including the non-equal retrieval budget, the author-constructed benchmark, low conflict-resolution scores, and unvalidated coefficients.

Auditable design with provenance. All operators are deterministic for a fixed configuration, and consolidated procedures keep the identifiers of the records they replace. For personal-assistant memory, inspectable provenance is a practical property that many LLM-driven memory pipelines lack, and the audit artifacts in Table 8 are potentially reusable.

Critical weaknesses

Confounded comparison and missing control experiments. Melo uses top-k = 20 (expandable to 60 on CMLM-Assist), while Vanilla RAG uses top-k = 5 and the external systems keep their native settings. The main tables therefore cannot separate Melo's memory mechanisms from a 4-12x larger evidence context. Several baseline numbers also look under-configured: MemGPT/Letta gets 9.5-12.1 F1 on LoCoMo, MemoryBank gets 1.34 F1 on CMLM-Assist (DeepSeek), and Mem0 on LoCoMo with GPT-4o-mini (28.54) seems well below what the original Mem0 paper reports. In the LoCoMo GPT-4o-mini block, Vanilla RAG, Mem0, A-Mem, and MemoryOS are almost identical on every metric, which suggests they may be getting near-identical contexts. Moreover, the matched-ablation control reaches 77.00 F1 while the main-table GPT-4o-mini result is 46.59, so configuration alone moves F1 by about 30 points. The following controls are needed: Vanilla RAG (and at least one memory baseline) at top-k = 20 and at a matched token budget, with the same backend and prompt as Melo. Baselines under the matched-ablation protocol (TF-IDF, top-k = 20, same prompt) where Melo reaches 77.00. A Full Context reference on LoCoMo. This is consequential because, without these controls, the main empirical claim cannot be attributed to Melo's memory design rather than to budget and configuration.

The closed-loop mechanism is largely inactive where it is evaluated. On LoCoMo, source records are mapped to protected episodic records, the deliberation trigger and retrieval gate are disabled, and there are no declared correction or task-outcome events for Eq. (6) to admit. Feedback admission, consolidation, and retention therefore seem mostly inert on the only independent benchmark, and the LoCoMo gains are more likely explained by retrieval scoring and budget. On CMLM-Assist, the direct test of the return path is the "w/o Feedback-to-Memory Update" ablation, but its diagnostic (procedures formed 96 -> 0) holds by construction. Conflict-resolution F1 stays below 8% on every backbone, and per-type results for temporal update, the case that motivates the paper, are not reported. The evaluation is also offline replay, so the loop never responds to Melo's own outputs. This is consequential because the closed loop is the main contribution, while the experiments mostly show the effect of profile-aware retrieval.

Oracle-like metadata in CMLM-Assist. In CMLM-Assist, importance, confidence, noise tags, and semantic tags are generated from the scenario specification, and Melo's rules consume exactly these fields (noise penalty in Eq. (9), tag routing in Eq. (11) including a benchmark-specific "university" tag, noise removal in Eq. (14)-(16)). Appendix D.6 shows that noise records score 0-0.083 and valid records score 0.586-0.878, so the distractors are removed by labels rather than by any inferred filtering. Baselines presumably do not use this metadata. In addition, all coefficients are hand-set on a development split that shares templates with the test split, with no sensitivity study. This is consequential because the largest margins (e.g., 46.59 vs. 20.58 F1) come from the benchmark whose construction matches Melo's input schema, while a real host would have to produce this metadata itself.

Limited novelty and overstated cognitive framing. The authors acknowledge that no individual operator is novel, and each is a simple rule (set-membership trigger, count-based consolidation, fixed exponential decay, keyword routing). Event-triggered revision, provenance, and procedural skill formation already appear in Reflexion, Voyager, Mem0, and A-Mem, so the claimed distinction (trigger, timing, provenance) is narrow. Prediction error and reconsolidation are central in the title and method, but the implementation uses none of their computational content (no error magnitude, no lability window). It is unclear what is conceptually new beyond an auditable rule-based memory pipeline.

Questions for authors

What are the results when Vanilla RAG and at least one memory baseline are run at top-k = 20 and at a matched token budget, and under the matched-ablation protocol where Melo reaches 77.00 F1? Does the same improvement trend hold?

On LoCoMo, which Melo components are actually active? How many feedback events are admitted, how many procedures are formed, and how many records are removed by retention? Could the authors provide a LoCoMo ablation?

How does Melo perform on CMLM-Assist when the importance, confidence, noise, and semantic tags are removed or inferred from raw text by an LLM? Conversely, how do the baselines perform when given the same metadata?

Could the authors report per-type results for all eight CMLM-Assist question types, especially temporal update and negative abstention?

Why are several baselines far below Vanilla RAG and below their published numbers, and why are the four GPT-4o-mini LoCoMo baselines nearly identical? Did any baseline fall back to a default path or receive empty memory?

Only lexical metrics are reported, and Melo's answers are about 2.5x longer than the references (12.43 vs. 4.95 tokens). Could the authors add an LLM-as-judge evaluation on LoCoMo and a length-controlled analysis?

How sensitive are the results to the hand-set coefficients in Eq. (7)-(9) and (13)-(15)? Also, what are the per-query latency, number of LLM calls, and context tokens for Melo compared with the baselines?

Reviewer_3WSaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents Melo, a cognitive science-inspired closed-loop memory framework for large language model agents. According to the paper, the framework integrates a rich set of mechanisms. Experiments are conducted on LoCoMo and the newly constructed CMLM-Assist benchmark dataset. Although maintaining and updating the long-term memory of agents is a critical problem, I believe this work suffers from severe issues in writing and methodological clarity, making it difficult to understand the improvements proposed in the paper.

Critical strengths

This paper explores the issues related to the maintenance and updating of long-term memory in large language model agents, and also conducts extensive evaluation experiments on multiple memory systems and large language model bases.

Critical weaknesses

W1. Poor Presentation and Structural Organization Obscure the Core Methodology. The manuscript suffers from a disorganized flow and lacks logical coherence, making it difficult to follow. Notably, there is a significant misalignment between the overview figure (Figure 1) and the main text, leaving several depicted modules unexplained. For instance, while Line 194 states that the framework comprises five main components, Figure 1 displays only four color-coded groups or nine individual blocks, making it difficult to establish a clear correspondence. Additionally, key terms are introduced prematurely without proper context (e.g., "Melo" first appears in Line 061 but is not formally introduced until Line 066).

W2. Limited Technical Novelty Compared to Existing Memory Updating Methods. The proposed framework primarily integrates existing techniques from the memory research domain and lacks a fundamentally novel mechanism that distinguishes it from established persistent memory systems such as Mem0, A-Mem, and MemoryOS. Furthermore, it remains unclear how the high-level cognitive science concepts discussed in the paper are concretely translated into operational computational mechanisms.

W3. Insufficient Mathematical Formalization and Implementation Details. Several core modules lack rigorous definitions, explicit assumptions, and clear explanations of inter-module interactions. For example, Equation (6) is formulated in a overly complex manner, yet the paper fails to clearly define the specific notations/symbols involved or explain how they are concretely computed.

W4. Questionable Experimental Evaluation and Lack of In-Depth Ablation Analysis. According to the reported evaluation, a simple Vanilla RAG baseline achieves the second-best performance, trailing only the proposed method. This raises questions about where the advantages of other specialized memory baselines lie and why these purposefully designed memory systems perform so poorly (often worse than Vanilla RAG). The paper lacks a thorough discussion and corresponding ablation studies to analyze this phenomenon—which is essential to clarify which specific design modifications in the proposed method actually drive the reported performance gains.

Questions for authors

What are the differences between Melo and other memory-related research works? Does it only perform integration, or does it propose improvement schemes combined with specific cognitive science mechanisms? In which part are the specific cognitive science mechanisms improved?

Reviewer_ix3DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Melo, a new framework for agent memory that uses an external memory component to manage, organize, and revise memories. Evaluated on an existing dataset and the newly constructed CMLM-Assist dataset, the proposed method achieves better performance on the evaluated tasks.

Critical strengths
  1. The motivation for this study is good. The proposed framework includes modules addressing different requirements and has a clear structure.

  2. To provide a more comprehensive evaluation, the authors use an existing dataset and construct a new dataset to assess the method’s effectiveness.

Critical weaknesses
  1. The main contribution of this paper appears to lie in workflow design. However, the framework contains several hyperparameters that seem to require task-specific tuning and optimization, such as those in Equations 7 and 8. Its applicability across different scenarios therefore remains uncertain.

  2. There lacks of recent baseline models that proposed in this year, as the existing baselines appear to have been published in 2025 or earlier. Given the considerable attention currently devoted to agent memory, omitting recent methods raises concerns about whether the comparison reflects the current state of the field.

Questions for authors
  1. Experimental settings should be explained more carefully. For example, in Section 5.1.4, why are the default retrieval sizes set to 20 for Melo and 5 for Vanilla RAG?

  2. The authors construct CMLM-Assist to evaluate capabilities such as feedback-based updating and preference retention. Although the paper acknowledges several limitations, would it be more appropriate to include human validation of the generated samples if the dataset is intended to serve as a new benchmark?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.