acceptodds
Under review as a conference paper at ICLR 2027

Preview-Guided Memory: Learning to Reread for Long-Context Reasoning

Abstract

LLM agents accumulate long contexts as they read documents, use tools, and carry out multi-step tasks. Memory-based methods handle these contexts by reading text in chunks and maintaining a compact memory. However, single-pass readers must decide what to retain before seeing later chunks, while still processing every chunk regardless of its relevance. We propose Preview-Guided Memory (PGM), which uses a preliminary preview to guide both reading and memory updates. In the first pass, PGM generates a short, question-independent preview of each chunk and groups neighboring previews into pages. A learned selector uses these pages and the question to choose which chunks to read in detail in the second pass. The reader then processes the selected chunks in order, using the corresponding preview page when updating its memory. These previews provide clues about upcoming content, while selection reduces the number of memory updates. We keep the reader fixed and train the selector with supporting-fact supervision followed by reinforcement learning, rewarding answer quality and evidence coverage, with a bonus for reduced reading when the answer is correct. PGM-Selective-7B achieved state-of-the-art on both HQA and longbench v2 with fewer tokens for reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
memory, long-context

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yEwSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends MemAgent-style chunked memory reading with a skim-then-read setup. It works in two passes.

  1. Preview: the model writes a short, question-independent preview of every chunk, and neighboring previews are grouped into pages.
  2. Select and read: a selector (a LoRA on the same 7B model) uses the previews and the question to pick which chunks to read. The reader then updates its memory over the selected chunks, with the neighboring previews in its prompt.

The authors find that PGM-Selective yields small improvements on HotpotQA (vs ReMemR1 and MemAgent) + on LongBench v2 (vs ReadAll).

Critical strengths
  • The paper is short and easy to follow, with a very clear Fig 1.
  • The problem is well-motivated. Single-pass memory readers have to decide what to keep before seeing what comes later.
  • Selection seems o really help at the longest lengths, and the paper does a good job counting the preview pass in its token costs.
Critical weaknesses
  1. The paper's main claim is that previews help the reader decide what to retain but the numbers don't fully support this. PGM-ReadAll, which has previews and reads everything, averages 75.00 on HQA, slightly below MemAgent (75.78), and is clearly worse at 448K (67.97 vs 72.66). The gains come from selection, which on HQA mostly means throwing away distractor documents.
  2. Since selection is doing the work, the key missing baseline is a cheap retriever choosing chunks at the same reading ratio (~45%), with the same reader: BM25 or dense top-k over raw chunks or over the previews. The random-chunks baseline (43.9/41.7) is too weak (and the DOS RAG baseline is unclear).

Smaller note: LycheeMemory's gate is a BCE chunk-relevance classifier trained on supporting facts with positive weight 3, which is essentially Eq. 5 here, so the related work undersells how close it is.

Questions for authors
  1. Can you run BM25 / dense chunk retrieval at matched reading ratio, feeding the same frozen PGM reader with and without previews in the update prompt?
  2. You have the memory text at every step, so the retention claim can be tested directly. Across all HQA questions, what fraction of gold supporting facts appear in the final memory for MemAgent, ReadAll, and Selective?
1 comment

Dear Reviewer yEwS,

Thank you for your careful review. We respond to your concerns and questions below.

W1: Evidence for preview-guided retention beyond the gains from selection.

A1: We thank the reviewer for this observation. We agree that our original presentation did not sufficiently distinguish the roles of previewing and selective reading. Selective reading indeed contributes substantially to accuracy by avoiding unnecessary memory updates, but its benefit is not merely discarding distractors. Our new controlled experiments provide evidence for both components. First, with the reader and selected chunks fixed, removing previews from memory updates reduces Selective accuracy by 2.34, 9.37, and 4.69 percentage points at 28K, 112K, and 896K, respectively, demonstrating that previews also influence what information is retained. Second, under approximately matched reading ratios and with the same PGM reader, our learned selector achieves 78.75% average HQA SubEM, compared with 72.66% for BM25 and 57.66% for dense retrieval. It also achieves substantially higher supporting-sentence recall. These results show that PGM combines contextual guidance for memory construction with learned, evidence-aware reading selection. We will revise the manuscript to clarify these distinct contributions.

W2 & Q1: Matched-reading retrieval baselines and preview ablations.

A2: We thank the reviewer for this suggestion. We conduct additional comparisons using BM25 and dense retrieval over raw chunks, with the same frozen PGM reader and approximately matched reading ratios. All retrieved chunks are restored to their original document order and processed using the same memory-update and answering procedure as PGM-Selective. The results on HQA are reported below (SubEM, %).

Method28K112K224K448K896KAverage
PGM-Selective82.8181.2576.5678.1375.0078.75
BM25-DOS + PGM Reader79.6964.8470.3175.7872.6672.66
Dense-DOS + PGM Reader69.5362.5055.4749.2251.5657.66

PGM-Selective outperforms BM25 and dense retrieval by 6.09 and 21.09 percentage points in average SubEM respectively, consistently across all five context lengths.

We further compare memory updates with and without previews across five HQA context lengths, keeping the reader and selected chunks fixed for each selection strategy.

SelectionWith PreviewWithout Preview
PGM Selector78.7576.09
BM2572.6669.84
Dense57.6655.78

Preview conditioning improves all three selection strategies. Moreover, PGM consistently outperforms BM25 and dense retrieval, demonstrating the benefits of both preview-guided memory updates and learned selection. We will include the complete results and the original DOS-RAG baselines details in the revised manuscript.

W3: Relationship to LycheeMemory’s supervised relevance gate.

A3: We thank the reviewer for pointing out this connection. We agree that our discussion of LycheeMemory was insufficiently detailed. The key distinction lies in how the two methods organize and use contextual information. LycheeMemory operates on compressed KV-cache representations and performs sequential, state-dependent gating based on the evolving working memory. In contrast, PGM constructs question-independent textual previews, groups neighboring previews into pages, and makes chunk-selection decisions before detailed reading. More importantly, PGM reuses these preview pages during memory updates, allowing the reader to leverage broader contextual cues when deciding what to retain. Furthermore, PGM optimizes the selector through trajectory-level RL with answer-quality and evidence-coverage feedback, while keeping the reader fixed. We will revise the Related Work section to better reflect the connection and clarify these distinctions.

Q2: Direct measurement of supporting-fact retention in final memories.

A2: We conduct a systematic audit of final memories from MemAgent, PGM-ReadAll, and PGM-Selective, measuring the fraction of gold supporting facts retained across all five HQA context lengths.

ContextMemAgentPGM-ReadAllPGM-Selective
28K67.49%82.66%80.80%
112K70.59%87.93%86.38%
224K66.87%86.69%86.38%
448K62.54%87.31%87.62%
896K60.06%86.69%86.69%
Average65.51%86.26%85.57%

PGM-ReadAll and PGM-Selective retain 20.75 and 20.06 percentage points more gold supporting facts than MemAgent on average, respectively. Notably, at 896K, both PGM variants retain 86.69% of gold supporting facts, compared with 60.06% for MemAgent. We will include the detailed analysis in the revised manuscript.

Reviewer_2YVZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Preview-Guided Memory (PGM) for long-context reasoning. It first generates question-independent previews of document chunks, then uses a selector trained with supervision and reinforcement learning to choose which chunks to read. A trained reader processes the selected chunks in source order, using previews to guide bounded-memory updates before answering. Experiments on HQA and LongBench v2 show improved accuracy over existing methods and reduced token use relative to the full-reading variant, PGM-ReadAll.

Critical strengths
  1. I find the problem of losing information before its relevance becomes apparent meaningful. Using previews for both reading selection and memory retention is an intuitive direction, and the theoretical construction and case study help explain its motivation.

  2. I appreciate the accounting of tokens across the complete inference pipeline and the explicit discussion of performance–cost trade-offs. PGM-Selective achieves promising overall results while substantially reducing token use relative to PGM-ReadAll, supporting the practical value of selective reading.

Critical weaknesses
  1. My main reservation is that the additional value of previewing is not sufficiently isolated. The experiments examine reinforcement learning and selection quality, but do not include a comparably trained sequential-memory baseline without previews or direct selection from raw text. On LongBench v2, ReadAll already achieves almost the same accuracy as Selective, leaving me uncertain how much of the improvement over MemAgent comes from preview conditioning, reader training, or other differences. More targeted controls would strengthen my confidence in the design.

  2. I would also appreciate clearer evidence of the benefits under comparable inference budgets. Relative to MemAgent, PGM-Selective improves accuracy but uses approximately 29% more tokens on HQA and 27% more on LongBench v2. The savings over ReadAll are useful, but do not establish an advantage over similarly costly alternatives. Its accuracy also matches MemAgent at 896K on HQA, motivating further analysis of the benefits and limitations at longer contexts.

Questions for authors
  1. I understand that preview pages provide compact cross-chunk context, but generating them also requires processing the entire source. Could the authors compare against selection directly from the question and raw text, including both individual chunks and neighboring chunks jointly? Reporting accuracy and complete token costs would help address my uncertainty about the added value of previewing.
  2. Could the authors include a sequential-memory baseline trained with comparable data and budgets but without previews? Additionally, holding the reader and selected chunks fixed while removing previews from memory updates would help distinguish the contributions of training, selection, and preview conditioning. These controls would be particularly helpful for understanding why ReadAll already approaches Selective’s accuracy on LongBench v2.
  3. The paper includes retrieval baselines, but their retrieval budgets and downstream answering procedures are not fully clear to me. Could the authors clarify these settings and consider feeding retrieved chunks through the same PGM memory-update reader? Comparisons under similar reading amounts or total token budgets would help establish the benefits over simpler retrieval-based alternatives.
  4. I would appreciate an analysis of possible information loss during preview generation. For example, do omitted entities or relations cause the selector to miss chunks containing annotated supporting facts? Distinguishing missing preview cues, selection errors, and information lost during memory updates would be helpful. Reporting evidence coverage across context lengths could also help explain the accuracy tie with MemAgent at 896K and clarify the method’s reliability on longer documents.
3 comments

Dear Reviewer 2YVZ,

Thank you for your careful review. We respond to your concerns and questions below.

W1: Isolating the contributions of previewing, reader training, and selection.

A1: We thank the reviewer for these helpful suggestions. To better isolate the contribution of previewing, we train a preview-free sequential-memory baseline using the same training data and budget as PGM. Across the five HQA context lengths, PGM-Selective achieves an average SubEM of 78.75%, outperforming the no-preview baseline (73.91%) by 4.84 percentage points. Moreover, when keeping the reader and selected chunks fixed, removing previews from memory updates reduces Selective's average accuracy by 2.66 percentage points, further demonstrating the contribution of preview conditioning. We also evaluate direct selection from raw text using two strategies: individual chunk selection and joint selection over neighboring chunks. Both use the same downstream PGM reader.

MethodAverage HQA SubEM
PGM-Selective78.75%
Raw Single-Chunk75.16%
Neighboring Chunks74.38%

PGM outperforms individual and neighboring raw-chunk selection by 3.59 and 4.37 percentage points, respectively. These results support the complementary benefits of preview-guided memory construction and learned selective reading. We will clarify these contributions in the revised manuscript.

W2: Accuracy–efficiency trade-offs under comparable inference budgets.

A2: We appreciate the reviewer's concern regarding the accuracy-efficiency trade-off under comparable inference budgets. To provide a more comprehensive analysis, we further conducted a per-query token-budget analysis across different HQA context lengths.

Single-query setting. Compared with MemAgent, PGM-Selective improves overall HQA accuracy. Moreover, the relative token overhead decreases progressively with increasing context length, demonstrating that PGM-Selective scales favorably to longer contexts while improving overall accuracy. These results suggest that the benefits of preview-guided selective reading become increasingly pronounced in longer-context settings.

Multi-query setting. Since the generated preview can be reused across multiple queries on the same document, we further analyze the amortized token cost, including the initial preview-generation cost. With 10 queries per document, PGM-Selective reduces average per-query token consumption by 7.0%–26.7% compared with MemAgent across all five HQA context lengths, while retaining the overall accuracy improvement. Even with four queries per document, it reduces average per-query token consumption by 4.9%–14.1% at context lengths of 112K–896K.

We will incorporate the corresponding accuracy-token analysis and multi-query efficiency results into the revised paper to provide a more comprehensive characterization of PGM-Selective's accuracy-efficiency trade-off.

Q1: Direct raw-text selection baselines and complete token costs.

A1: We further compare PGM with two preview-free raw-text selection strategies: individual chunk selection and joint selection over neighboring chunks.

MethodAverage HQA SubEMAverage Tokens
PGM-Selective78.75%665.44K
Raw Single-Chunk75.16%457.48K
Neighboring Chunks74.38%1,867.68K

Compared with individual raw-chunk selection, PGM achieves 3.59 percentage points higher accuracy with moderately increased token consumption. More importantly, compared with neighboring-chunk selection, PGM improves accuracy by 4.37 points while reducing token consumption by 64.4%. These results demonstrate that PGM provides a favorable accuracy-efficiency trade-off: it achieves better reasoning accuracy than direct raw-text selection while avoiding the substantial computational overhead of repeatedly processing neighboring raw chunks.

We will include these raw-text selection baselines and detailed results in the revised manuscript.

Q2: A comparably trained preview-free baseline and fixed-selection preview ablation.

A2: We thank the reviewer for these suggestions. We conduct two additional controlled experiments to examine the contributions of previewing and selective reading. First, we train a sequential-memory baseline without previews using the same training data and budget as PGM. Across the five HQA context lengths, PGM-Selective achieves an average SubEM of 78.75%, outperforming the no-preview baseline (73.91%) by 4.84 percentage points. Second, we remove previews from memory-update prompts while keeping the trained reader and selected chunks fixed. The results show that preview conditioning improves Selective's average HQA accuracy by 2.66 percentage points across the five context lengths, demonstrating its contribution beyond chunk selection. Together, these results demonstrate the advantage of combining preview-guided memory construction with learned selective reading.

We will include these experiments and results to distinguish the contributions of training and preview conditioning.

Q3: Retrieval baselines with comparable reading amounts and the same PGM reader.

A3: We conduct additional comparisons using BM25 and dense retrieval over raw chunks, with the same frozen PGM reader and approximately matched reading ratios. All retrieved chunks are processed using the same PGM memory-update and answering procedure.

Method28K112K224K448K896KAverage
PGM-Selective82.8181.2576.5678.1375.0078.75
BM25 + PGM Reader79.6964.8470.3175.7872.6672.66
Dense + PGM Reader69.5362.5055.4749.2251.5657.66

PGM-Selective achieves 6.09 and 21.09 percentage points higher average SubEM than BM25 and dense retrieval, respectively. We will include these results in the revised version.

Q4: Evidence coverage and information loss across the pipeline.

A4: We thank the reviewer for this suggestion. We conduct an evidence-coverage analysis and an error audit across all five HQA context lengths (28K–896K). PGM achieves 96.28%–98.45% supporting-sentence recall and 93.75%–97.66% complete evidence-chain coverage under limited reading budgets. Even at 896K, PGM maintains 97.83% supporting-sentence recall and 96.88% complete evidence-chain coverage. We also analyze errors across the pipeline, identifying 8 selection errors, 22 memory-update errors, and 33 final-answering errors across the five context lengths. We will incorporate detailed evidence-coverage and error-attribution figures into the revised manuscript to illustrate evidence preservation and information loss at different context lengths.

Thank you for the additional experiments. The fixed-reader ablation addresses my main concern about the contribution of previews on HQA. On balance, I still view the methodological advance as modest, so I will keep my rating at 2. I appreciate the care you took in addressing my questions.

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

This paper proposes Preview-Guided Memory, a two-pass reading framework where the reader first writes short question-independent previews of every chunk and groups them into pages. A learned selector then chooses which chunks to read in detail, and the reader updates a bounded memory on the selected chunks while seeing the matching preview page. With the selector trained by supporting-fact supervision and RL, PGM-Selective-7B reaches 78.75 average SubEM on HQA up to 896K tokens and 34.79% on LongBench-v2.

Critical strengths
  • The learned selector beats random selection by 37.34 points at the same 46% of tokens, showing that what is read matters more than how much.
  • Selective reading degrades only from 82.81 to 75.00 SubEM between 28K and 896K while full reading falls to 68.75, which is the practical case for the method.
Critical weaknesses
  • The paper credits its gain to preview-guided memory, but on HQA selection beats full reading by 10.16 points at 448K while ReadAll does not beat MemAgent, and LongBench v2 reverses this, so which component carries the gain remains open.
  • PGM assumes question-independent previews keep what matters later, but this moves the decide-before-knowing problem it attributes to single-pass readers into the preview stage, so a chunk whose preview omits the needed detail may never be selected.
Questions for authors

Clarification

  • PGM-ReadAll sees a preview at every update yet does not beat MemAgent on HQA, so what do previews contribute to memory, as opposed to deciding what to skip?
  • Previews are written without the question, which is the decide-before-knowing condition the paper criticizes in single-pass readers, so why should a short preview anticipate later relevance better than a memory update does?

Future work

  • The introduction describes context that grows as an agent acts, so what should a preview summarize when the later chunks do not yet exist?
4 comments

Dear Reviewer RXKW,

Thank you for your careful review. We respond to your concerns and questions below.

W1 & Q1: Contributions of preview-guided memory and selective reading.

A1: We thank the reviewer for this comment. We clarify that selective reading in PGM serves not only to reduce inference costs but also to improve answer accuracy. Reading all chunks introduces unnecessary memory updates, whereas selecting informative chunks helps preserve useful information and avoid interference from irrelevant content. Previewing provides additional cross-chunk context to guide these memory updates.

We provide two controlled experiments to substantiate the contributions of both components. First, using the same frozen PGM reader, approximately matched reading ratios, and identical preview-conditioned memory-update and answering procedures, PGM-Selective achieves an average HQA SubEM of 78.75%, substantially outperforming BM25 (72.66%) and dense retrieval (57.66%). This demonstrates that learned selection provides benefits beyond simply reducing the number of chunks processed. Second, keeping the reader and selected chunks fixed, removing previews from memory updates reduces PGM-Selective's average HQA SubEM by 2.66 percentage points across the five context lengths. Notably, at the longest context length of 896K, preview conditioning improves accuracy by 4.69 percentage points, highlighting the value of cross-chunk contextual information when reasoning over particularly long documents.

These results establish the complementary contributions of the two components: learned selection improves which information is read, while preview conditioning helps determine what information is retained. We will revise the manuscript to clarify the complementary roles of preview conditioning and selective reading and include these controlled comparisons.

W2: Potential information loss in question-independent previews.

A2: We agree that question-independent previews are inherently lossy and may omit details that later become relevant. Thus, PGM does not eliminate the risk of missing evidence; rather, it addresses a broader trade-off shared by long-context methods: sequential memory may discard facts during repeated memory updates, while selective reading may fail to retrieve chunks whose relevance is difficult to determine in isolation. The key distinction is that PGM does not require previews to preserve all answer-bearing details. Instead, previews provide coarse-grained topical and contextual cues that help identify potentially relevant chunks and guide memory updates. For example, in our Figure 2 case, the preview page exposes the topic of a later film entry without revealing the author relation or birth date. This contextual cue helps the reader retain the earlier birthday until the raw text of chunk 22 provides the missing author relation. Our objective is therefore not lossless reading, but to mitigate information loss under a bounded computation budget by providing a broader view of the document before detailed reading and memory updates. We will add these discussions in the revised version.

Q2: How question-independent previews help anticipate later relevance.

A2: We clarify that question-independent previews are not intended to predict future answers or explicitly identify which facts will be queried. Rather, they provide coarse-grained contextual cues about surrounding content, helping the reader avoid prematurely discarding potentially useful information. Our existing Hunter Davies example (Figure 2) illustrates this mechanism through delayed entity-relation discovery. We further provide a second example involving delayed geographical association (HQA 112K, question 50). An early chunk states that Newport is the county seat of Sullivan County. After an intervening discussion of other counties, MemAgent loses this association. Later chunks establish that East Lempster belongs to Sullivan County, but the necessary county-seat fact is no longer available. PGM retains Newport through the intervening update and correctly answers the question. Notably, the preview page available during this update already lists subsequent entries about Lempster and East Lempster, without revealing Newport or the complete geographical relation. Together, these examples illustrate how previews can provide useful contextual cues for preserving potentially relevant facts before their downstream relevance becomes explicit, without directly supplying the answer. We will add these analysis in the revised manuscript.

Q3: Preview generation for dynamically growing agent contexts.

A3: We thank the reviewer for this insightful suggestion. We agree that extending PGM to dynamically growing agent contexts is an interesting direction, and we will further explore incremental preview generation as new chunks become available in our future work.

Thank you for the detailed clarifications. Another questions raised that.

You explain that 'Readall' introduces unnecessary memory updates, while selection helps preserve useful information and avoid interference from irrelevant content. However in another response yEwS, Readall and selection has quite similar fraction of gold supporting facts, so selection does not seem to preserve more useful information. If irrelevant content can enter the memory without pushing out gold facts, how large is the memory budget? Since the budge is upper bound, could you also explain the average final memory length for ReadAll, and selection? If Readall has much more memory length, then it makes sense.

Thank you for raising this point. Both PGM-ReadAll and PGM-Selective use the same PGM reader with a fixed memory capacity of 1,024 tokens. Although their gold supporting-fact retention is comparable, Selective produces more compact final memories across all five context lengths.

The following table reports the mean, median, and P5–P95 range of final memory lengths (in tokens), calculated over 128 questions per context length.

ContextReadAll MeanSelective MeanReadAll MedianSelective MedianReadAll P5–P95Selective P5–P95
28K151.56143.26138129.554.40–278.7054.40–278.70
112K276.32239.6923821087.90–563.5081.35–488.50
224K317.39284.16296254106.10–682.1083.40–625.60
448K373.59331.03348.5315.5139.70–711.40112.45–589.10
896K434.42410.84380.5370.5148.75–875.30135.80–829.20
Average310.66281.80280.20255.9079.95–701.0072.95–645.50

Selective reduces the average final memory length by 9.29%. Its median memory length is also lower at every context length.

Furthermore, Selective reduces the irrelevant-information ratio across all five context lengths, with an average reduction of 3.24 percentage points compared with ReadAll. It also reduces sequential memory-update operations by approximately 55% at 112K–896K, while achieving higher average HQA accuracy.

These results show that selective reading maintains comparable gold-fact retention while producing more compact memories with less irrelevant information and fewer updates. We will incorporate these detailed statistics into the revised manuscript. We sincerely hope these additional results and clarifications address your concerns. Thank you again for your constructive feedback.

Thank you for your detailed responses and additional analyses. I appreciate your efforts to address my concerns. However, I will maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

update pdf

Dear Reviewers,

We have submitted a revised version of our manuscript, incorporating the additional experiments, analyses, and clarifications discussed in our rebuttal. We have also updated the paper to better clarify the respective contributions of previewing and selective reading.

Thank you again for your constructive feedback and valuable suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.