acceptodds
Under review as a conference paper at ICLR 2027

CERES: RECONCILING DURABLE EVIDENCE AND EVOLVING ACTIVE STATE FOR LONG-LIVED LLM AGENTS

Abstract

Large language models (LLMs) are increasingly used to build long-lived agents, which rely on external memory to support their long-term operation. However, the core challenge of memory goes beyond retrieval: a system must preserve raw interaction evidence before memory compression. It must also revise existing records and preserve the basis for each revision as facts change, reconcile conflicting records, and prevent outdated conclusions from continuing to dominate agent behavior. Many memory methods emphasize retrieval, while preserving information during writing, updating facts over time, and resolving conflicts remain challenging. Omitted evidence may be irrecoverable downstream, contradictions may persist, and higher-order conclusions synthesized from interaction histories may become outdated without active maintenance. We introduce CERES (Continual Evidence Reconciliation and Evolving State), which unifies these operations as an evidence-grounded memory lifecycle. CERES comprises two core components. Engram first preserves raw interaction evidence, then extracts recall-worthy information into structured records and integrates these records into structured memory pages with source links. Cortex maintains a capacity-limited set of higher-order representations grounded in these pages. It admits, refreshes, or retires entries based on changes in their supporting evidence, recorded usage and feedback, making them available as default-active context. Across end-to-end comparisons on Memora and an auxiliary evaluation on LongMemEval-S, CERES demonstrates more reliable long-term memory performance. Write and recall experiments further show that Engram consolidates cross-session evidence into reliable memory that is traceable and aligned with the current state. Under identical memory content, comparing answering with and without Cortex guidance shows that it improves average answer quality.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model agents, long-term memory, memory systems, provenance-aware memory, evolving memory, memory retrieval

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

Long-term memory is not only about whether historical information can be retrieved. Original evidence may be lost during compression, while outdated abstract evidence may still affect agents. CERES builds an evidence-based memory system. Engram preserves full evidence and integrates memory, while Cortex produces high-level conclusions. The main experiments show that CERES improves over Memora.

Critical strengths

The research question is clearly defined. It concerns how to preserve evidence, store facts, and maintain conclusions in long-term memory. It combines both the engram and cortex parts.

Critical weaknesses
  • The number of samples in LongMemEval-S is small (60 tasks?). Essentially, the main experiments are conducted on only one dataset.
  • Abstraction + evidence is not a new idea. Could the authors further explain the novelty of the proposed method?
  • It feels weird, possibly AI-generated, to capitalize many English words, e.g., Monthly, Consolidate, Trace, and Judge.
Questions for authors
  • The experimental setup is not clearly described. For example, readers cannot know what reasoning effort/settings are used for the LLMs. It's also confusing to denote Qwen3.7-Max as Qwen3.7. GPT-6-Astra is used but never mentioned in the table or its caption.
  • Why is there blue text used in Figure 4?
Reviewer_vDq5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The goal is to use two components to maintain durable "state" for LLM agents. One extracts recall-worthy information, and another exctracts representations based on this. There's a sort of systems building component to manage (admit / refresh / retire etc.) these components.

Critical strengths

Seems to target some relevant critical components of LLMs with long-term state.

Experiments are not bad, though somewhat mixed. The somewhat mixed results may limit enthusiasm for the paper given other issues.

Even if the results are a bit mixed, the analysis and various comparisons are fairly thorough.

Critical weaknesses

Very light on technical details. These days (where most papers are written with LLMs), it's hard to trust papers like this with little exposition of the technical details and a more "narrative-heavy" structure.

It's hard to spot a specific intellectual contribution to get excited about. This seems more a like a systems-building type of approach, gluing reasonable enough parts together to build something that might help a bit in some contexts. But I think without a more exciting core intellectual contribution, it's hard to picture authors following up on this work. The bar for intellectual contribution should (I think) be fairly hard to pass for ICLR these days.

Experiments are (I think) against some standard (though reasonable enough) baselines. Given the mixed results, I have some doubt as to how this paper compares to SotA techniques. This would perhaps be partially addressed in a response.

Questions for authors

See above. Useful to answer queries especially w.r.t. comparison against SotA.

Reviewer_XHCQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of external memory for long-lived agents and proposes CERES, a framework that organizes memory capabilities into an evidence-anchored memory lifecycle. Engram first preserves raw interaction evidence in an append-only trace; after Organize and Reconcile, this evidence is consolidated into structured memory pages with provenance links and retained revision history. Cortex maintains a capacity-limited set of evidence-anchored high-level conclusions as the default context, and dynamically admits, refreshes, or evicts them as supporting evidence, usage, and feedback change. The framework targets three stages: write integrity, fact reconciliation, and active-state evolution.

Critical strengths
  • S1. Clear problem formulation and convincing motivation: Memory is decomposed into a complete lifecycle covering write integrity, fact reconciliation, and active-state maintenance. The two failure scenarios in Figure 1 are concrete and realistic.
  • S2. Coherent architecture that separates evidence from state: The append-only trace ensures that no evidence is lost, while structured pages jointly retain the current state, revision history, and provenance.
  • S3. Complete write-side ablation with targeted metrics: Stale Leakage, Update Tracking, and Current-State F1 directly measure current-state accuracy and stale-history handling.
  • S4. Significant context-efficiency advantage: Memory-Context Tokens show a clear advantage over MEMORA and Nemori.
Critical weaknesses
  • W1 (Novelty): Each write-side component has a direct counterpart in prior work, as the paper acknowledges in Lines 83–93. The append-only evidence layer corresponds to APEX-MEM; the separation between raw evidence and derived semantics corresponds to Zep; write-time reconciliation corresponds to Mem0's add/update/delete, A-Mem's link updates, and Nemori's NEW/MERGE/CONFLICT; and the combination of current state, revision history, and provenance corresponds to Zep's bi-temporal validity/invalidity-time mechanism and THEANINE's temporal chain. The genuine increment therefore narrows to the specific design that co-locates three states within a memory page and the evidence-gating criteria in Cortex. Moreover, the paper dismisses the closely related Zep with only “separates raw episodes from the semantic facts” and “introduce temporal structure,” without discussing how Zep's bi-temporal invalidation mechanism differs from the reconciliation mechanism proposed here.
  • W2 (Baselines): The main experiments compare only against Mem0, MEMORA, Nemori, while several closest methods are not included as baselines. Consequently, the core gap claim—“jointly preserving traceable factual revisions and keeping high-priority conclusions aligned with accumulating evidence remains challenging”—is not tested against these baselines, nor is it grounded in an analysis of the specific failure modes under which existing methods cannot jointly achieve these goals; it remains at the level of an assertion.
  • W3 (Experimental model choice): Table 12 in the appendix shows that the Consolidate stage is fairly sensitive to the model, yet the main experiments report only results for the GLM-5.2 + Sonnet-4.6 combination. In addition, the main experiments specifically adopt a two-model division of labor for evaluation, so it cannot be ruled out that the advantages stem from the benefits of model division of labor rather than from the architecture itself.
  • W4 (No ablation of the heuristic function): The matching score formula between query qq and entry hh in E.1 is a heuristic function with many manually tuned hyperparameters, and no sensitivity analysis is provided.
  • W5 (Component coupling): The existing ablations demonstrate that individual components are useful (e.g., Reconcile helps, structured pages help, and the Cortex read strategy helps), but they do not isolate the contribution of the coupling mechanism whereby revision of an Engram page triggers Cortex re-evaluation.
Questions for authors
  • Can the authors provide additional analysis of why and how CERES is effective in solving the core problem of “jointly preserving traceable factual revisions and keeping high-priority conclusions aligned with accumulating evidence”? In particular, please add an ablation that isolates the coupling mechanism (e.g., fixing Cortex so that it does not retire/refresh when an Engram page is revised) and observe the degree of degradation in metrics such as Stale Leakage. Please also provide specific and accurate comparisons with the closest baselines to clarify the gap, and add end-to-end results for Zep and A-Mem on MEMORA and LongMemEval-S.
  • It would be helpful to provide a detailed table of model configurations for each stage of each system, and to add a single-model end-to-end run of CERES compared with the baselines.
  • The Mem0 results seem abnormally low, especially on LongMemEval-S. Could the authors provide the specific configuration?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.