acceptodds
Under review as a conference paper at ICLR 2027

ConteXtNext: Learning to Predict Context Needs as Knowledge Evolves

Abstract

External memory enables LLM agents to accumulate experience across tasks without updating model parameters. Yet stored experience does not automatically become effective context. Turning accumulated knowledge into useful context requires learning from past usage to guide knowledge selection and composition for subsequent tasks. Knowledge evolution introduces a further difficulty: revisions change the content available for selection, while the feedback guiding selection remains tied to earlier versions. The challenge is to improve the stored knowledge while preserving and updating what has been learned about its use. We introduce ConteXtNext (CXN), a framework for learning to select and allocate context as its knowledge base evolves. CXN decouples free-form knowledge organization from utility estimation, preserving traceable historical evidence across knowledge evolution and reconstructing utility on demand without maintaining persistent, task-agnostic scores. It formulates context construction as budgeted maximization of task-conditioned set utility. Task rewards and retrospective self-assessment provide approximate contribution credit, while coverage evidence accounts for overlap within candidate contexts. These estimates guide the selection of complementary knowledge through predicted marginal gains in set utility, with the initial context budget adapted to the information demands of related past tasks. We evaluate CXN across three task domains: AppWorld, CRMArena-Pro, and KernelBench. Across the evaluated LLM configurations, CXN improves task success over no-memory baselines by up to 18.7, 37.5, and 11.0 percentage points, respectively. Further analysis on KernelBench shows that CXN's performance gains are accompanied by lower context consumption. CXN uses the fewest context tokens among the compared methods, consuming only 25.2% of the no-memory baseline's context tokens while increasing task success from 74% to 85%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Memory, Context Prediction, Knowledge Evolution

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_mW4gOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ConteXtNext (CXN), a framework for constructing task-specific context from an evolving external memory for LLM agents. Its three main contributions are:

1.Instead of attaching static scores (e.g., Q-values) directly to memory files, CXN stores historical usage observations in a separate evidence store and tracks explicit transformation lineage (across splits, merges, and revisions) under conservation constraints so historical credit transfers without duplication.

  1. It formulates context construction as budgeted set-utility maximization: reconstructing item utility on demand by weighting similar past tasks, combining task rewards with retrospective self-assessment, discounting overlapping memories via directional coverage edges, and dynamically scaling the initial context budget to task demand.

  2. Evaluated on AppWorld, CRMArena-Pro, and KernelBench with two LLM backbones, CXN improves task success over no-memory baselines (by up to 18.7, 37.5, and 11.0 percentage points) and outperforms memory baselines in 4 of 6 settings while significantly reducing context token/character consumption.

Critical strengths
  1. The paper identifies a practical limitation in existing agent memory systems, that attaching static Q-values or importance scores directly to memory units breaks when memories are reorganized (split, merged, or revised). Separating the free-form editable Markdown workspace from an immutable, version-bound evidence store is a well-motivated systems design that supports both flexible knowledge refactoring and historical credit assignment.
  2. CXN is evaluated under a strict online streaming protocol (without future leakage) across three distinct domains (AppWorld, CRMArena-Pro, and KernelBench) and two LLM backbones. It consistently improves over no-memory baselines (up to +37.5 pp on CRMArena-Pro and +18.7 pp on AppWorld) and outperforms four recent memory baselines (ACE, ReasoningBank, ReMe, MemQ) in 4 of 6 settings, while substantially reducing main-agent initial context size (-68.8% vs. BM25) and context token consumption (-74.8% vs. No Memory on KernelBench).
Critical weaknesses
  1. The paper does not isolate its main methodological contributions: conservation-constrained lineage (τv\tau_v), coverage-based redundancy discounting (co′oc_{o'o}), and adaptive budgeting (B(q,v)B(q,v)). The only retrieval ablation (FTS recall) replaces all three simultaneously with BM25 and only drops performance by 4% on AppWorld (76% to 72%). Moreover, the paper never reports how often splits, merges, or revisions actually occur during the experiments.
  2. Table 2 claims a 75% reduction in token usage, but only counts main-agent context tokens. It omits the LLM token overhead from CXN's background memory operations (retrospective self-assessment, Manager revisions/splits/merges, and LLM reranking), even though Section 4.1 states that total tokens including memory operations were tracked.
  3. The equations rely on continuous scalars (τv(b,a)∈[0,1]\tau_v(b,a) \in [0,1] for split inheritance and co′o∈[0,1]c_{o'o} \in [0,1] for directional coverage) estimated by prompting an LLM. The paper neither evaluates the sensitivity or calibration of these LLM-generated numbers nor provides the prompts, Manager schemas, or exact lexical supplementation formula needed to reproduce the system.
Questions for authors
  1. How frequently do knowledge transformations (split, merge, revise, archive) actually occur across the AppWorld, CRMArena-Pro, and KernelBench streams? Furthermore, can you provide ablations that isolate the individual impact of (a) transformation lineage conservation (τv\tau_v), (b) directional coverage discounting (co′oc_{o'o}), and (c) adaptive budgeting (B(q,v)B(q,v)), rather than collapsing all three into the FTS recall baseline?
  2. Section 4.1 notes that "total logged model tokens including memory operations" were tracked, whereas Table 2 only reports main-agent context tokens. What is the total token consumption and wall-clock/API cost of CXN compared to the baselines once background memory operations (retrospective self-assessment, Manager revisions/splits/merges, and LLM reranking) are included?
  3. How are the continuous split inheritance shares τv(b,a)∈[0,1]\tau_v(b,a) \in [0,1] and directional coverage values co′o∈[0,1]c_{o'o} \in [0,1] elicited from the LLM, and how sensitive is CXN to noise or miscalibration in these estimates (e.g., compared to uniform 1/k1/k split shares or binary coverage)? Additionally, could you provide the exact formula used to combine lexical relevance with set-utility gain Δv(a∣q,S)\Delta_v(a \mid q, S) during greedy selection?
Reviewer_FmxFOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ConteXtNext (CXN), a framework for context construction from evolving agent knowledge. The key idea is to separate free-form knowledge from historical utility evidence, and use transformation lineage to preserve the applicability of evidence when knowledge is revised, split, or merged. For new tasks, CXN reconstructs task- and set-conditioned utility from historical evidence and selects complementary knowledge under an adaptive context budget. Experiments on AppWorld, CRMArena-Pro, and KernelBench show improved task success and reduced context consumption in several settings.

Critical strengths
  1. Meaningful problem. The paper identifies an interesting issue: historical utility feedback may become misaligned when stored knowledge evolves.

  2. Reasonable formulation. Separating knowledge from historical evidence and using lineage with conservation constraints provides an intuitive way to handle knowledge transformations. Set-conditioned utility and Coverage are better motivated than simply ranking individual memory items independently.

Critical weaknesses
  1. The central contribution is not directly validated. The main novelty is evidence continuity under knowledge evolution, but there are no controlled comparisons against variants such as CXN without lineage, resetting evidence after knowledge evolution, or naive utility inheritance. Therefore, the experiments mainly demonstrate the effectiveness of the full system rather than the proposed evidence-continuity mechanism.

  2. The experimental analysis is insufficient relative to the complexity of the framework. The ablation study only examines two coarse variants on a 100-task subset of AppWorld, leaving key components such as lineage/conservation, Coverage, and adaptive budgeting insufficiently isolated. The reliability of semantic correspondence and evidence transfer is also not evaluated. Moreover, CXN does not outperform the strongest baseline on KernelBench (85% vs. 89% with MiMo, and 99% vs. 100% with DeepSeek), but the paper provides limited analysis of why the proposed mechanism is less effective in this domain. Understanding this failure case would help clarify the applicability and limitations of CXN.

Questions for authors
  1. Could the authors directly compare CXN with no lineage/conservation, naive evidence inheritance, and evidence reset after knowledge evolution?
  2. How does performance change as the frequency or severity of revision, split, and merge operations increases?
  3. How accurate are the semantic correspondence, Usage, and Coverage estimates, and how sensitive is CXN to errors in these quantities?
  4. Could the authors provide stronger component ablations, multiple runs/task orders, and end-to-end computational costs?
Reviewer_FuohOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ConteXtNext (CXN) addresses context construction for LLM agents whose external knowledge evolves over time. It separates free-form knowledge organization from structured historical evidence, tracks lineage across knowledge revisions, splits, and merges, and reconstructs task- and set-conditioned utility from historical usage, coverage, and reward signals to guide context selection under an adaptive budget. The authors evaluate CXN on AppWorld, CRMArena-Pro, and KernelBench, reporting improvements over no-memory baselines in several settings and reduced context consumption. The problem is important, and the framework provides a coherent approach to evidence-aware context construction. However, the current evaluation primarily establishes end-to-end system performance; it does not sufficiently establish the semantic validity of evidence transfer or the predictive validity and independent contribution of the proposed utility-estimation mechanisms.

Critical strengths
  1. Important and timely problem. The paper addresses a meaningful challenge in long-running agents: historical feedback may become difficult to interpret as external knowledge is revised and reorganized. Separating knowledge organization from evidence storage is a coherent design choice.

  2. A technically structured framework with broad evaluation. Version lineage, conservation-constrained support transfer, set-conditioned utility, and adaptive budgeting form a plausible integrated approach. Evaluation across three task domains, including context-consumption analysis, provides useful end-to-end evidence that the overall system can help in some settings.

Critical weaknesses
  1. The claimed continuity of historical evidence across knowledge evolution is not sufficiently established. CXN decouples knowledge organization from utility estimation by retaining version-bound historical observations and transferring their support through explicit transformation lineage. However, the conservation constraint guarantees only that inherited support is not numerically duplicated; it does not establish that the transferred evidence remains semantically applicable after revisions, splits, or merges. In particular, semantic inheritance shares are supplied by the Knowledge Manager, and the paper does not sufficiently validate whether these assignments preserve the applicability of historical evidence. Consequently, the paper's central claim that historical evidence remains applicable as knowledge evolves is not adequately substantiated.

  2. The effectiveness and necessity of the proposed utility-estimation and context-selection mechanisms remain unclear. CXN combines task similarity, reward-guided contribution estimates, coverage correction, set-conditioned utility, and adaptive budgeting to select context. Although the end-to-end results show improvements in several settings, the evaluation does not directly establish whether predicted utility gains reflect the actual contributions of selected knowledge. Furthermore, the reported ablations do not isolate the contributions of the main mechanisms or establish their advantages over simpler relevance- and redundancy-aware retrieval strategies. Thus, the effectiveness of the proposed estimator and the necessity of its added complexity remain insufficiently demonstrated.

Questions for authors
  1. How are the numerical signals used by CXN actually obtained? I am not fully clear on how the key inputs to the utility estimator are instantiated in practice. Appendix B mentions lexical matching, task embeddings, and optional LLM reranking for historical task matching, while Coverage coefficients and semantic inheritance shares are described at a higher level. Could the authors clarify the concrete procedures used to compute task similarity (s(q,t)), assign Coverage coefficients, map task rewards, and determine inheritance shares? In particular, how does the Knowledge Manager determine the semantic inheritance shares, and which parts of this process are automated versus supplied through explicit judgments?

  2. What exactly does the predicted set utility measure, and how should it be interpreted? The paper defines utility through approximate contribution targets derived from historical usage, task similarity, Coverage, and reward signals, but I find it difficult to determine how closely this quantity is intended to reflect the actual contribution of a knowledge unit to task performance. Is the estimator primarily intended to rank candidate contexts, or to estimate their actual task-dependent utility? How do the authors establish that a higher predicted marginal gain generally corresponds to a more useful context in practice, rather than merely reflecting the assumptions built into the estimator?

  3. How much does each component contribute to the reported improvements? The current ablations remove task-reward feedback or replace the entire evidence-guided selection pipeline with fixed-budget BM25 retrieval, making it difficult to distinguish the effects of lineage-based evidence transfer, Coverage correction, set-conditioned utility estimation, and adaptive budgeting. Why were these mechanisms not evaluated separately? In particular, how much of the improvement can be attributed to the proposed set-utility formulation and adaptive budget, rather than to historical task matching or simpler relevance- and redundancy-aware retrieval?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.