acceptodds
Under review as a conference paper at ICLR 2027

Prediction Is Not Protection: What Models Forget During Context Compaction

Abstract

Long-running LLM agents replace old conversations with compact memory. Protecting selected facts can improve overall recall while making information left unprotected less reliable. We study this tension through a paired evaluation that links each fact's pre-compaction forecast, untreated omission, and outcome after key-only protection. With supplied importance labels and controlled fact budgets, gains in overall important-fact recall coexist with losses on the same unprotected facts under self-selected, random, and cue-based policies. At the widest memory-loop budget, external protection retains 76–93% of raw-log recall at 10–29% of its storage, but a simple repetition rule remains competitive; self-ranking is model-dependent. Naturalistic tests qualify these findings. LoCoMo establishes neither general cue superiority nor a best selector. On 12 Multi-Session Chat dialogues at approximately matched measured storage, total point gains occur without detectable unprotected-fact loss. Additional key-free controls do not establish a benefit from stronger wording; self-selection retains a nominal gain over plain compaction. These limited-source results do not establish the absence of collateral cost or a universal selector advantage. Prediction accuracy, overall recall, and reliability on the same unprotected facts therefore require separate evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, context compaction, agent memory, metamemory, self-prediction, memory systems

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

This paper studies if LLMs can predict facts that will be lost during history compaction and whether the predictions can help choose facts to explicitly preserve in the compaction summary. Fact protection experiments were conducted on 3 open-weights models and show that LLM generated predictions do not provide a clear preservation gain over other methods such as random choice or a simple repetition rule. The authors also find that explicit requests to preserve important facts can improve overall recall but lead to reduced recall of other facts. More realistic settings, LoCoMo and 12 Multi-Session Chat dialogues, however do not clearly support the same findings, which were established in the synthetic settings.

Critical strengths
  • The paper includes appropriate controls such as tracking the same set of unprotected facts across different settings and include comparisons to random and repetition baselines.
Critical weaknesses
  • Clarity: The paper is difficult to follow. Nearly all claims in the paper are heavily qualified and includes many key terms that are not clearly defined (depth_frac, outcome rate, interference cue, common-valid) or rely on inconsistent naming (repeat cue, repetition rule, mention count all refer to the same thing).
  • Significance: The main findings of the paper are inconclusive. The main fact protection experiment is only conducted on three small open-weight models. LLM self-prediction is a weak policy and the authors do not identify any selection policy that is reliably best. Further, the key finding that overall recall hides the reduced recall of unprotected facts does not hold in more natural/realistic settings. The paper does not make clear why its specific settings and experiments matter and findings do not provide a clear takeaway for the reader.
Questions for authors
  1. What is the main takeaway a reader should draw from this paper? It appears the main goal is to measure if prediction is a useful policy for improving fact recall under constrained compaction settings, but this is not communicated clearly.
  2. Why should the reduced recall of unprotected facts not be taken as an expected resulted of the limited budgets that models are evaluated under? The compaction settings are capped to 3-9 facts, which seems detached from realistic agent compactions which typically span thousands of tokens.
Reviewer_LCS5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

LLM agent cannot carry the entire history and need to compact the conversations into memory. The paper discussed the question: protect some facts comes at the expense of the reliability of others. The paper proposed some experiments to verify this question and has certain empirical experiment results to proof this question.

Critical strengths

None. I cannot find the evidence to support the acceptance of this paper.

Critical weaknesses

The biggest concern for this paper is the contribution. I can see the authors find a hypothesis, and design a set of experiments to proof this hypothesis. and then? I cannot find any useful information about how we use this hypothesis and conclusion to resolve the important problem in the research community or industrial. The hypothesis studied in this paper is not a counterintuitive one. even without these empirical experiments, most people would believe the hypothesis by common sense. usually the process of compacting the conversation history is a trade-off of performance and LLM capacity.

Questions for authors

see the weakness.

Reviewer_NAU5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether protecting selected facts during LLM context compaction hurts the facts left unprotected. It proposes a fact-level evaluation that compares recall before and after protection on the same set of unprotected facts. On synthetic work logs with fixed fact budgets, protection improves overall recall but lowers recall on unprotected facts, and models cannot reliably predict what they will forget. On naturalistic data, the results are inconclusive on LoCoMo, and no unprotected-fact loss is detected on Multi-Session Chat. The authors present the main contribution as an evaluation workflow rather than a new method.

Critical strengths

The paper builds a fairly complete experimental chain, from controlled synthetic logs to a memory loop and then to naturalistic datasets (LoCoMo and Multi-Session Chat), which shows which findings hold under control and which do not carry over.

Critical weaknesses
  1. The paper is very hard to follow. Many new concepts are introduced without explanation. For example, in the introduction, it is unclear what "reliable" means in "making the same unprotected facts less reliable". The second paragraph is overly compressed, and terms such as omission prediction, forecast, untreated summary, keys, decision value, and selector are not defined at all. Similar problems continue throughout the paper, e.g., what are "cached outputs"? What does "matched storage" mean? As a result, I spent a long time on the paper but could not understand what it was saying.
  2. The contribution is unclear. This is partly due to the writing, but the authors also acknowledge that the collateral-loss result does not extend to the naturalistic datasets, and that the contribution is not a universal protection rule or a proven mechanism, but only an evaluation workflow. As a result, the paper has no clear core conclusion, and it does not align well with the title. "What models forget during context compaction": the paper does not clearly answer what models forget.
  3. The finding on synthetic data, that protecting some facts decreases the recall of unprotected facts, is likely caused by the memory budget constraint and is quite intuitive, which undermines this finding. Moreover, this finding does not generalize to LoCoMo and MSC, possibly because their constraints are looser.
Questions for authors
  1. Can the authors show that the collateral loss is not simply a consequence of the fact budget?
  2. Why does the collateral loss disappear on MSC?
  3. Could the authors state the main takeaway of the paper in one sentence? Given that the collateral-loss finding does not carry over to LoCoMo and MSC, and that the contribution is described as an evaluation workflow rather than a method or mechanism, it is unclear what new knowledge the paper offers to the community.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.