acceptodds
Under review as a conference paper at ICLR 2027

COVER: Causally Ordered, Verified Event Records for Schedule-Robust LLM Memory

Abstract

Stateful language-model memory can interpret the same event differently depending on which other events have already arrived. Deduplicating or sorting records after extraction cannot repair this semantic change. We introduce COVER, which compiles each event from its causal ancestors, retains source-bound records, and joins them independently of delivery order. Under authenticated causal dependencies and a fixed deterministic compiler, COVER produces identical terminal states and answers across legal schedules and retries. Our analysis characterizes the local conditions for this guarantee and the information that compression must preserve for future queries. Experiments separate structural reliability from semantic utility. In a controlled action relay, COVER improves success across all tested schedules by 2.50 percentage points over typed FIFO with the same content and budget. Source-retaining readers also improve complete-history LongMemEval and LoCoMo accuracy in selected configurations. Strong canonical controls account for much of the reliability benefit, however, and restrictive typed adapters lose utility on natural dialogue. COVER provides a causal interface for reliable memory construction; its semantic gains depend on the compiler, reader, and evidence budget.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM memory, causal consistency, schedule robustness, retrieval-augmented generation, language model agents

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ejv3OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies when LLM memory produces the same behavior under reordered and retried event delivery. COVER interprets immutable events using their exact causal ancestors, validates extracted records, preserves unresolved source material, and uses deterministic merging and readout. The theoretical analysis separates schedule invariance from semantic correctness.

Critical strengths

The most useful contribution is the separation between interpretation and storage: a convergent merge cannot repair an event interpretation that already depended on an arbitrary concurrent arrival. The causal-ancestor interface makes the required assumptions explicit, including fixed compilation and trustworthy dependency information. This is a meaningful systems-design contribution even where the underlying algebra is classical.

Critical weaknesses
  • The empirical need for ancestor-restricted interpretation is less directly demonstrated than the general need for deterministic handling of delivery. Exact-ID deduplication already removes all correctness flips in the retry experiment, and Appendix H’s cached canonical control matches COVER’s record maps and compilation lookups. These are appropriate controls, but leave less direct evidence about how often interpreting an event against concurrent, non-ancestor records causes a consequential failure in realistic traces.

  • The learned representation does not yet have a consistently beneficial operating configuration. In the separate official-code comparison in Section 5.3, typed COVER averages 26.12 versus 44.20 for raw sources; the two-call cascade also loses to Raw-2 for Llama in Appendix K. The paper acknowledges this, so I view it as a limitation on practical significance rather than a contradiction of the convergence guarantees. Total compilation and readout cost is also not matched by equalizing reader-call limits.

Questions for authors

Which experiment compares ancestor-conditioned compilation with current-store-conditioned compilation while keeping the model, events, merge rule, and readout fixed? If there is no such comparison, which empirical result supports the specific benefit of excluding non-ancestor context?

Where do the parent links come from in each evaluation? Which results assume externally supplied causal metadata, and which test dependencies inferred from the observations?

Are replicas reusing the same compiled records or independently invoking the LLM? In the latter case, what enforces the fixed-compiler assumption required by the convergence result?

Do the diagnostics identify where the accuracy loss in Section 5.3 occurs: compilation, retention, retrieval, or interpretation by the reader?

What are the total token and model-call costs of the facts/raw cascade relative to Raw-2, including amortized compilation? Equal reader-call limits do not establish matched end-to-end cost.

Reviewer_ar1aOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COVER, a framework for schedule-robust LLM memory construction. The method compiles each event using its causal ancestors, retains source-bound records, and combines them using order-independent operations. Under assumptions of accurate causal dependencies and deterministic compilation, the framework guarantees consistent memory states across different event arrival orders.

The authors provide theoretical characterizations of schedule robustness and evaluate the framework on synthetic event streams, downstream action execution, and several long-term memory benchmarks.

Critical strengths
  1. The paper identifies a meaningful consistency issue in stateful LLM memory systems: different event arrival orders can lead to different interpretations and memory states.

  2. The theoretical analysis provides a formal characterization of schedule robustness, and the experiments distinguish structural consistency from downstream semantic utility.

Critical weaknesses

1. The central causal-dependency assumption is not adequately justified in realistic LLM-agent settings.

The proposed method fundamentally relies on the runtime supplying accurate causal dependencies between events. However, the paper does not provide a mechanism for obtaining or validating these dependencies in realistic LLM-agent interactions.

In natural-language conversations and asynchronous agent workflows, event dependencies are frequently implicit, ambiguous, or distributed across multiple interactions. It is unclear how an external runtime can reliably determine which previous events a new message semantically depends on.

The authors explicitly acknowledge that COVER cannot infer missing dependencies from timestamps or textual similarity. Consequently, the central guarantee is conditional on precisely the information that may be difficult to obtain in practical deployments.

This is a major limitation of the problem formulation. The paper establishes that memory can be made order-independent when the correct causal structure is already available, but does not demonstrate how this prerequisite can be satisfied in realistic applications.

2. Limited methodological novelty and insufficient practical benefit over existing approaches.

The main components of COVER—causal dependency tracking, immutable event records, exact-ID deduplication, deterministic compilation, and associative, commutative, and idempotent merging—are largely established concepts from distributed systems and event-sourcing architectures.

While their integration into LLM memory is potentially useful, the paper does not sufficiently isolate a substantial new algorithmic contribution beyond applying these existing principles.

Furthermore, the paper reports that a cached canonical insertion baseline achieves identical structural behavior with the same number of compilation lookups. This raises questions about what unique capability COVER provides beyond a carefully implemented conventional consistency mechanism.

The downstream evaluation also fails to establish a compelling practical advantage. In the official-code comparison across five static benchmarks, COVER achieves 26.12% average accuracy compared with 44.20% for raw retrieval. The improvements on sequential tasks are not statistically significant.

Overall, the paper provides a conditional consistency guarantee, but does not convincingly demonstrate either sufficient methodological novelty or substantial utility improvements to justify acceptance.

Questions for authors
  1. How are the causal dependencies required by COVER obtained in real-world LLM-agent systems? In particular, when dependencies are implicit in natural-language interactions rather than explicitly tracked by the runtime, how does the proposed framework identify the correct causal ancestors?

  2. What happens when the supplied causal graph is incomplete or incorrect? Can the authors quantify how COVER's consistency and semantic accuracy degrade under missing, spurious, or incorrectly specified dependencies? This seems essential for assessing its practical applicability.

  3. What is the key algorithmic novelty beyond existing causal-consistency and event-sourcing mechanisms? Given that the cached canonical baseline achieves equivalent structural behavior and compilation lookup counts, which capability or efficiency improvement is unique to COVER?

  4. What evidence supports deploying COVER instead of simpler canonical raw-memory approaches? The official-code results show a substantial accuracy deficit relative to raw retrieval, while the downstream execution improvements over strong canonical controls are limited.

Reviewer_5XFLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper intended to be a memory paper that allowed to process information in a causal manner to react appropriately to requests that might be affected by earlier requests in the past.

Critical strengths

Paper is so difficult to read I don't think there is any benefit from doing so

Critical weaknesses

The paper was very heavy (if not in its entirety) writing by AI. I tried to read the introduction many times and still couldn't even get from the manuscript what was the high level Intuition behind the paper, a clear problem description and what is being done in the paper.

The text os clearly structured as an AI output, sentences that look legit English but don't really connect and don't say the information that should be in a scientific paper. I could not understand how the proposal works and even less how it was evaluated in the empirical validation.

Questions for authors

N/a

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.