Artificial Intelligence · arXiv:2608.21690 v1 · submitted 21 Aug 2026 · 77 references
Context as an Environment: Programmatic Context Management for Long-Horizon Agents
The paper says
“LLM agents increasingly take on long-running tasks whose history grows far beyond a single model context window. Existing approaches compress earlier interactions or extract selected information into fixed memory representations, committing to what to preserve before future needs are known.”
more of the abstract
We present Scroll, a context manager that treats each agent session as an executable Session Environment. The environment is backed by an append-only Event Log and a sandboxed, persistent Python kernel. The kernel maintains a typed namespace across model calls, allowing tool outputs, retrieved history, and derived state to be bound to variables rather than serialized into the prompt at each call. Model-written code searches, materializes, and transforms session state through exec; only explicitly printed projections enter the model's working view for the next call. Context management thus becomes a programming task that inherits the improving coding abilities of LLMs, while the Event Log preserves lossless historical ground truth. As the working view approaches its budget, stale spans are evicted but remain recoverable: an eviction index keeps compact landmarks tied to exact Event Log addresses, so that the agent navigates directly to evicted regions instead of searching the full log. With Qwen3.8-Max as the backbone, Scroll achieves 94.8% on LongMemEval_S; 73.1% on BEAM_10M, surpassing the best published memory system by 5.1 points; and 86.7% on LOCA_256K, exceeding the best published long-horizon agent by 37.4 points.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 27 Aug 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A well-built context-management system whose internal ablations support the mechanism, but the headline benchmark margins over published baselines are not established under matched conditions. This line is wrong
4 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 77 references say.
The paper says, at Abstract; Tables 2 and 7
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The headline comparative claims are not established by the evidence, and the paper's own captions disclaim the comparisons. The abstract states Scroll 'surpasses the best published memory system by 5.1 points' on BEAM 10M (73.1 vs 68.0 for Exabase M-1) and 'exceeds the best published long-horizon agent by 37.4 points' on LOCA 256K (86.7 vs 49.3 for MiniMax M3 + ReAct).
more of objection 1
However, Table 2's caption says the rows are 'reference points from the literature rather than a controlled comparison: reader models differ across rows and can substantially affect scores,' and Table 7's caption notes that prior systems use different backbones. Exabase M-1 uses Gemini 3 Flash as its reader, and MiniMax M3 is a different model family used with ReAct; neither is re-run here. The 5.1-point and 37.4-point margins therefore conflate method differences with backbone, reader-model, judge, and evaluation-configuration differences. This is load-bearing because the abstract's SOTA claims are the paper's primary public framing. The authors should either (a) temper the abstract and conclusion to 'competitive under each system's own configuration' with the uncontrolled nature stated, or (b) add controlled comparisons by running at least one open baseline (e.g., Mem0 or Hindsight) with the same Qwen3.8-Max backbone, the same Qwen3.6-flash judge, and the same protocol as Scroll. The footnote in Section 4.1 that baselines are not reproduced because independent reproductions have led to disagreements explains a practical choice but does not resolve the confound.
The paper says, at Section 3.3 (Evaluation Protocol)
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
Each task is evaluated once with a random seed, and no variance estimates or confidence intervals are reported. This is a serious limitation for the comparative claims. In particular, the LOCA 256K difference between Scroll (86.7) and the CodeAct agent (85.3) is only 1.4 points, yet Section 4.1 presents Scroll as achieving the best performance in a controlled comparison; a single stochastic rollout cannot support a 1.4-point margin as meaningful.
more of objection 2
The BEAM 10M headline margin of 5.1 points likewise has no error bar, and BEAM uses an LLM-as-a-judge which adds judge-side variability (mitigated here by temperature 0, but the agent trajectories are still stochastic). The paper states in Section 4.3 that the ablation profile is 'consistent across our repeated runs,' which indicates that repeated runs exist and should be reported with means, standard deviations, and ideally per-seed results. At minimum, the authors should run 3–5 seeds for the headline tasks (LongMemEval S, BEAM 10M, LOCA 128K/256K) and report the full distribution, and should hedge any conclusion whose margin is smaller than the observed spread.
The paper says, at Section 3.2 and Appendix C (Benchmark-specific rubrics)
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The benchmark-specific rubrics give the agent task-specific schema knowledge and protocol guidance that amounts to a prompting advantage. Each memory benchmark contributes a rubric that specifies the exact data layout (kind strings such as 'beam chat turn' and 'context msg', the seq column, session id, created_at), the memory-surface API (ms.search, ms.expand, ms.sql_query, ms.days_between), and even the exact answer/abstention phrasing (e.g., "I don't have that information from our conversations" for LongMemEval).
more of objection 3
The paper claims 'no few-shot demonstrations,' which is literally true, but the rubric is a compact, benchmark-specific instruction set telling the agent how to query each benchmark's data schema and how to phrase responses. This reduces the external-validity claim that Scroll is a general context-management mechanism: the comparisons against literature baselines are not matched for prompt information, and the method's strong scores may partly reflect rubric engineering rather than the core mechanism. The internal comparisons (Table 3, Figure 3) hold the rubric fixed and are therefore unaffected. The authors should either ablate the rubric (e.g., provide a generic memory-surface description without the benchmark's kind strings, column names, and exact abstention phrasing) or explicitly scope the claims as benchmark-configuration-dependent.
The paper says, at Table 6 (Per-category BEAM 10M results)
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The per-category breakdown undercuts the blanket 5.1-point claim and shows the aggregate margin is not a uniform capability gain. Scroll trails Exabase M-1 substantially on summarization (70.5 vs 91.9) and temporal reasoning (47.5 vs 58.8), and trails Hindsight on preference following (89.1 vs 97.5).
more of objection 4
The overall lead is driven by a few categories where Scroll is very strong—knowledge update (92.5 vs 45.0–75.0), contradiction resolution (88.1 vs 32.5–58.8), and information extraction (75.0 vs 51.2–66.3)—which is a coherent mechanism-level story (ordered verbatim evidence beats write-time digests). But because the baseline category scores come from different reader models and judges, even the category-level leads are not verified. The paper's own caption says cross-system comparison is 'indicative rather than controlled.' The Section 4.1 sentence 'Scroll leads by the widest margins where the answer hinges on a few exact records' is an interpretation, not a demonstrated result. The authors should either verify the category-level differences against at least one open baseline run under the same judge and backbone, or consistently describe all cross-system numbers as uncontrolled reference points.
6 minor objections
The paper says, at Throughout
Pith says: minor objection 1 of 6
Several typesetting artifacts place missing spaces around inline monospace identifiers: 'We presentScroll', 'throughexec', 'anexecutableSession Environment', 'theagent harness', and 'a working view over a persistent Session Environment' in the contributions list. Please fix these in the final PDF/HTML.
The paper says, at Section 2.4 (Eviction index complexity)
Pith says: minor objection 2 of 6
The claim that the eviction index occupies O(k log_k n) blocks after n evictions is stated without derivation or citation. Since k is the tier width, the dependence on k should be made precise (O(k log_k n) = O((k/log k)·log n)), and a short proof sketch or reference should be provided. If this is an asymptotic bound over worst-case eviction patterns, state the assumed pattern.
show 4 more
The paper says, at Figure 3 caption
Pith says: minor objection 3 of 6
The caption reports 'mean judge score' with hatched bars, but Section 3.3 says each task is evaluated once. Please clarify whether the plotted values are means over the questions within the single run, and if multiple runs exist, report them or add error bars.
The paper says, at Table 4 and Section 4.2
Pith says: minor objection 4 of 6
Kimi-K2.7 scores 30.7 at LOCA 128K and 32.0 at LOCA 256K, a non-monotonic pattern worth a footnote; the current text says the spread 'widens to 64 points,' but the non-monotonicity suggests high variance, which should be acknowledged.
The paper says, at Appendix D and rubrics
Pith says: minor objection 5 of 6
The term 'step_index' is used interchangeably with 'session id' across the rubrics and trajectory code blocks (e.g., Code Block 2 prints 'step_index' while the BEAM rubric says 'each session has a distinct session id'). Use one term consistently.
The paper says, at Section 4.4
Pith says: minor objection 6 of 6
The statement that BEAM 10M's median input of 105K tokens is 'about 1% of the corpus' depends on the exact corpus size; please state the corpus token count or write '≈1%' to avoid rounding confusion.
What would settle it
Re-run Exabase M-1 (or the best published BEAM 10M system) with the Qwen3.8-Max backbone, the same retrieval budget, and the same Qwen3.6-flash temperature-0 judge used for Scroll, on the same BEAM 10M split; if the re-run score exceeds 68.0 or comes within noise of Scroll's 73.1, the claimed 5.1-point lead is not established. Separately, run the MiniMax M3 + ReAct baseline on LOCA 256K with the same judge and environment description used for Scroll, and report whether 49.3 reproduces; if it rises substantially, the 37.4-point claim needs revision.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 27 Aug 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The core idea is genuinely good: keep the full session in a persistent, executable environment (Event Log + Python kernel) and let the model write code to decide what enters its context, rather than compressing or extracting at write time. The integration of existing pieces — CodeAct, programmatic tool calls, event-sourced logs — into a lossless session environment with recoverable eviction is new enough, and the paper is honest about what it builds on. The detailed appendix trajectories, including two failures with post-hoc analysis, are a real strength; you can see exactly where the mechanism works and where query formulation or sampling fails.
What holds up: Table 3 fixes the backbone and toolset and shows Scroll and CodeAct beating summarization and retrieval agents with much smaller degradation from 128K to 256K. Figure 3 shows lossy summarization collapses (19.9 vs. 73.1), removing the persistent REPL costs 7.3 points, and removing the eviction index costs only 1.8 — modest but still positive. The cross-backbone sweep is also useful, showing the method transfers but depends on model coding ability. Cost numbers (median 105K input tokens on BEAM 10M, about 1% of the corpus) support the practical claim.
The soft spot is exactly where the stress-test note lands: the abstract's headline margins over prior systems are not established. Table 2's caption itself says the baselines are reference points with different reader models and judges, and Table 7 uses different backbones. The 5.1-point lead over Exabase on BEAM 10M is fragile — Exabase beats Scroll on summarization (91.9 vs. 70.5) and temporal reasoning (58.8 vs. 47.5), so the overall margin mixes method gains with backbone and judge differences. The 37.4-point lead on LOCA compares against systems with different backbones and strategies, so it's not a controlled win. Also, each task is run once with a single seed; no error bars anywhere. A minor circularity: the benchmark-specific rubrics in the prompts encode knowledge of each benchmark's data layout, though they look like generic search guidance rather than answer keys.
My take: the mechanism survives scrutiny — the controlled internal comparisons are solid and the failure analysis is refreshingly honest. But the abstract overclaims. The authors should either reproduce the key baselines under matched conditions or reframe the literature comparison as indicative and lead with Table 3 and Figure 3.
This paper deserves a serious referee. It's a well-constructed system with a plausible design, transparent evaluation of its own components, and actionable if imperfect benchmarks. Send it to review, but insist on matched-baseline evidence or tempered claims, plus variance reporting. I'd cite it if I worked on context management, and I'd bring it to reading group.
The strongest objection
The abstract's central comparative claims—BEAM 10M +5.1 points over the best published memory system (73.1 vs. 68.0 for Exabase M-1) and LOCA 256K +37.4 points over the best published long-horizon agent (86.7 vs. 49.3 for MiniMax M3 + ReAct)—depend on the comparability of literature baselines that the paper itself disclaims. The Table 2 caption states these are 'reference points from the literature rather than a controlled comparison: reader models differ across rows and can substantially affect scores.' Table 7 uses different backbone models. Since Exabase M-1 and MiniMax M3 are proprietary and not re-run here, the headline margins conflate method gains with backbone, reader-model, judge, and evaluation-configuration differences. The 5.1-point lead over Exabase M-1 on BEAM 10M is especially fragile: BEAM is judged by an LLM-as-a-judge, and Exabase's score may use a different judge or reader; the paper's own Table 6 shows per-category scores where Exabase beats Scroll on summarization (91.9 vs. 70.5) and temporal reasoning (58.8 vs. 47.5), so the overall margin is not a uniform capability gain. Within the paper, the controlled comparisons are much more secure: Table 3 fixes the Qwen3.8-Max backbone and toolset across four context-management strategies, and the ablation in Figure 3 is internally consistent and supports the lossless-event-log and persistent-kernel mechanism. The LOCA 128K vs. 256K drop (-2.6) versus CodeAct's (-4.0) is a controlled, internally valid signal. So the load-bearing weakness is not the mechanism but the published-margin claims; those specific numbers should be downgraded to 'not established under matched conditions' unless the baselines are reproduced.
The reader's weakest_assumption was that the literature baselines are not controlled comparisons and therefore the headline leads may not hold. I agree with this and identify it as the most load-bearing concern: the abstract's signature numbers depend on cross-system comparability that the paper explicitly disclaims. The paper's internal controlled comparisons (Table 3, Figure 3, Table 4) support the mechanism's efficacy and are properly designed, so the central contribution is not undermined. I recommend keeping the CONDITIONAL verdict: the system and mechanism claims are credible and partially supported, but the headline margins over prior systems should be restated as uncontrolled reference points until matched-condition reproductions exist. The concrete test would settle whether the specific margins survive matched evaluation. I do not see a fatal internal inconsistency; the concern is external comparability, which the paper itself flags. The baselines are proprietary, so the limitation is in the evidence, not the authors' conduct.
The referee's report
If the mechanism holds, Scroll is a meaningful contribution to long-horizon context management: it shifts from write-time compression/extraction to query-time programmable construction over lossless state, a principled and well-motivated design. The controlled elements are genuinely strong: Table 3 fixes backbone and toolset across four context-management strategies; Figure 3 ablates the lossless event log, the persistent kernel, and the eviction index; and Appendix D publishes real trajectories, including failures, which lets a reader see exactly where the method succeeds and where it falls short. The reproducibility link and the explicit acknowledgment of the fragility of the preference-following and summarization categories are honest and useful. The per-category results in Table 6 reveal a coherent capability profile: Scroll wins where exact ordered evidence matters (knowledge update, contradiction resolution) and loses where write-time digesting helps (summarization, preference following). This is a falsifiable, mechanism-level claim that the ablations and trajectories substantiate.
Pith's reading of the paper (a machine's summary, not the authors')
Scroll claims that an LLM agent's accumulated history should not be serialized into the prompt at every step; the history should live outside the context as a persistent Session Environment that the model programs against. The paper reports 94.8% on LongMemEval S, 73.1% on BEAM 10M, and 86.7% on LOCA 256K with its strongest tested backbone, surpassing published prior systems by 5.1 points on BEAM 10M and by 37.4 points on LOCA 256K. A sympathetic reader should care because, if the claims hold, context management stops being a lossy summarization problem and becomes a code-synthesis problem, inheriting the rapid progress of LLMs at writing programs.
What carries the argument. The load-bearing mechanism is the Session Environment: an append-only Event Log for ground truth, an external payload store for large tool outputs, and a sandboxed persistent Python kernel with a typed resident namespace. The model-facing surface is a memory object that factorizes context construction into four operations: search (deterministic full-text retrieval that returns stable sequence addresses), materialize (expand an address or address range back into exact events and payloads), compute (ordinary Python plus permitted SQL, filesystem, and tool interfaces over resident variables), and expose (print, the only boundary across which a projection enters the model's context). A third component, the eviction index, keeps a compact landmark bound to each evicted span's sequence range and rolls older landmarks into coarser tiers, so the working view stays bounded while every removed span remains recoverable.
Core claim. The paper's central claim is that context management is a programming task. Scroll records every event of an agent session in an append-only Event Log with stable addresses and keeps tool outputs and derived values as typed variables in a persistent Python kernel; the model writes code that searches the log, materializes spans, and computes over resident objects, and only what the model explicitly prints becomes the next model call's working view. No history is lost: when the working view grows past budget, evicted spans are folded into compact, address-anchored index entries, so the agent can navigate back to exact event ranges instead of guessing what to ask for. This flips the question 'what should the model see next?' from a decision made when events arrive to a program the model writes at query time.
The premise it rests on. The headline margins assume the published scores of prior systems are comparable even though those runs used different reader models, backbones, and judges; the paper itself flags this in the caption to Table 2.
If the paper is right (Pith's inferences, not the authors')
If the claims hold, the bottleneck for long-horizon agents shifts from how much context to pack into a window to how well the model can write retrieval, joining, and aggregation code.
Production agents that currently compact or summarize the prompt could instead keep a lossless event log plus an eviction index, trading prompt-token budget for lightweight code execution.
The reported token costs suggest a fixed context budget can serve histories far larger than the window: on BEAM 10M, the median input working view is about 105K tokens for a 10M-token corpus, because only printed projections enter context.
Because the context-management policy is code, stronger backbones should benefit more on tasks requiring multi-step query planning, so further gains should arrive with better code-writing models and no change to the harness.
The same query-time mechanism should also let an agent answer questions whose evidence is spread across many sessions by composing small programs rather than by hoping a retrieval pipeline anticipated the question.
Editorial extension: if context management tracks code synthesis, then distilling successful retrieval-and-query programs into smaller models should raise their scores most on aggregation-heavy benchmarks such as LOCA 256K, where the paper already finds the largest backbone gap.
Editorial extension: the lossless-log-plus-anchored-index pattern could be reused beyond agents, in audit trails, reproducible data pipelines, or debugging tools where a bounded view over an unbounded history must remain reversibly navigable.
Editorial extension: the reported margins over prior systems are directional until a matched-control evaluation fixes the same backbone, reader model, and judge across all compared systems; the paper itself identifies this caveat for its literature-sourced baseline table.
Editorial extension: one could make the eviction index adaptive by tracking which evicted regions are actually revisited and keeping finer landmarks there, rather than a fixed tier width.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
LLM agents increasingly take on long-running tasks whose history grows far beyond a single model context window. Existing approaches compress earlier interactions or extract selected information into fixed memory representations, committing to what to preserve before future needs are known. We present Scroll, a context manager that treats each agent session as an executable Session Environment. The environment is backed by an append-only Event Log and a sandboxed, persistent Python kernel. The kernel maintains a typed namespace across model calls, allowing tool outputs, retrieved history, and derived state to be bound to variables rather than serialized into the prompt at each call. Model-written code searches, materializes, and transforms session state through exec; only explicitly printed projections enter the model's working view for the next call. Context management thus becomes a programming task that inherits the improving coding abilities of LLMs, while the Event Log preserves lossless historical ground truth. As the working view approaches its budget, stale spans are evicted but remain recoverable: an eviction index keeps compact landmarks tied to exact Event Log addresses, so that the agent navigates directly to evicted regions instead of searching the full log. With Qwen3.8-Max as the backbone, Scroll achieves 94.8% on LongMemEval_S; 73.1% on BEAM_10M, surpassing the best published memory system by 5.1 points; and 86.7% on LOCA_256K, exceeding the best published long-horizon agent by 37.4 points.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.84
“With Qwen3.8-Max as the backbone, Scroll achieves 94.8% on LongMemEval_S; 73.1% on BEAM_10M, surpassing the best published memory system by 5.1 points; and 86.7% on LOCA_256K, exceeding the best published long-horizon agent by 37.4 points.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.20 · human participants 0.04 · controlled comparison 0.61 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.05
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| LLM agents increasingly take on long-running tasks whose history grows far beyond a single model context window. | background 1.00 | 0.05 | 0.02 | asserted 0.99 |
| Existing approaches compress earlier interactions or extract selected information into fixed memory representations, committing to what to preserve before future needs are known. | background 0.66 | 0.03 | 0.03 | asserted 1.00 |
| We present Scroll, a context manager that treats each agent session as an executable Session Environment. | contribution 0.99 | 0.87 | 0.12 | asserted 1.00 |
| The environment is backed by an append-only Event Log and a sandboxed, persistent Python kernel. | method 0.96 | 0.76 | 0.04 | asserted 1.00 |
| The kernel maintains a typed namespace across model calls, allowing tool outputs, retrieved history, and derived state to be bound to variables rather than serialized into the prompt at each call. | method 0.98 | 0.76 | 0.05 | asserted 1.00 |
| Model-written code searches, materializes, and transforms session state through exec; only explicitly printed projections enter the model's working view for the next call. | method 0.99 | 0.71 | 0.09 | asserted 1.00 |
| Context management thus becomes a programming task that inherits the improving coding abilities of LLMs, while the Event Log preserves lossless historical ground truth. | interpretation 0.75 | 0.54 | 0.41 | asserted 1.00 |
| As the working view approaches its budget, stale spans are evicted but remain recoverable: an eviction index keeps compact landmarks tied to exact Event Log addresses, so that the agent navigates directly to evicted regions instead of searching the full log. | method 0.99 | 0.89 | 0.08 | asserted 1.00 |
| With Qwen3.8-Max as the backbone, Scroll achieves 94.8% on LongMemEval_S; 73.1% on BEAM_10M, surpassing the best published memory system by 5.1 points; and 86.7% on LOCA_256K, exceeding the best published long-horizon agent by 37.4 points. | result 1.00 | 0.96 | 0.41 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 7,057 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (4)
Figures are the authors', reproduced from the paper.
References (77)
- [1] Loca-bench: Benchmarking language agents under controllable and extreme context growth , author=. arXiv preprint arXiv:2602.07962 , year= on PitharXiv
- [2] Tavakoli, Hamidreza and others , journal =. Benchmarking and Enhancing Long-Term Memory in on Pith
- [3] Acon: Optimizing context compression for long-horizon llm agents , author=. arXiv preprint arXiv:2510.00615 , year= on PithPith's reviewarXiv
- [4] Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents , author =. arXiv preprint arXiv:2601.01885 , year = on PithPith's reviewarXiv
- [5] Zhou, Zijian and Qu, Ao and Wu, Zhaoxuan and Kim, Sunghwan and Prakash, Alok and Rus, Daniela and Low, Bryan Kian Hsiang and Liang, Paul Pu , booktitle = on Pith
- [6] Wei, Tianxin and Sachdeva, Noveen and Coleman, Benjamin and He, Zhankui and Bei, Yuanchen and Ning, Xuying and Ai, Mengting and Li, Yunzhe and He, Jingrui and Chi, Ed H. and Wang, Chi and Chen, Shuo and Pereira, Fernando and Kang, Wang-Cheng and Cheng, Derek Zhiyuan , journal = on Pith
- [7] Froger, Romain and Andrews, Pierre and Bettini, Matteo and Budhiraja, Amar and Cabral, Ricardo Silveira and Do, Virginie and Garreau, Emilien and Gaya, Jean-Baptiste and Lauren. arXiv preprint arXiv:2509.17158 , year = on PitharXiv
- [8] Motwani, Sumeet Ramesh and Nichols, Daniel and London, Charles and Li, Peggy and Pizzati, Fabio and Blake, Acer and Hammoud, Hasan and McDonald, Tavish and Naik, Akshat and Ivanova, Alesia and Baskaran, Vignesh and Laptev, Ivan and Glatt, Ruben and Ben-Nun, Tal and Torr, Phili... on Pith
- [9] Recursive Language Models , author =. arXiv preprint arXiv:2512.24601 , year = on PithPith's reviewarXiv
- [10] Ehrlich, Clint and Blackman, Theodore , journal =. 2026 , howpublished = on Pith
- [11] Code Execution with. 2025 , howpublished = on Pith
- [12] Wu, Di and Wang, Hongwei and Yu, Wenhao and Zhang, Yuwei and Chang, Kai-Wei and Yu, Dong , booktitle = on Pith
- [13] Hu, Yuanzhe and Wang, Yu and McAuley, Julian , journal =. Evaluating Memory in on Pith
- [14] Modarressi, Ali and Deilamsalehy, Hanieh and Dernoncourt, Franck and Bui, Trung and Rossi, Ryan A. and Yoon, Seunghyun and Sch. International Conference on Machine Learning (ICML) , year = on Pith
- [15] Kuratov, Yuri and Bulatov, Aydar and Anokhin, Petr and Rodkin, Ivan and Sorokin, Dmitry and Sorokin, Artyom and Burtsev, Mikhail , booktitle = on Pith
- [16] Laban, Philippe and Hayashi, Hiroaki and Zhou, Yingbo and Neville, Jennifer , journal = on Pith
- [17] Packer, Charles and Fang, Vivian and Patil, Shishir G. and Lin, Kevin and Wooders, Sarah and Gonzalez, Joseph E. , journal = on Pith
- [18] Shang, Jingbo and Zheng, Zai and Wei, Jiale and Ying, Xiang and Tao, Felix , journal = on Pith
- [19] Behrouz, Ali and Li, Zeman and Kacham, Praneeth and Daliri, Majid and Deng, Yuan and Zhong, Peilin and Razaviyayn, Meisam and Mirrokni, Vahab , journal = on Pith
- [20] The Markovian Thinker: Architecture-Agnostic Linear Scaling of Reasoning , author =. arXiv preprint arXiv:2510.06557 , year = on PitharXiv
show 57 more
- [21] Agentfold: Long-horizon web agents with proactive context management , author=. arXiv preprint arXiv:2510.24699 , year= on PitharXiv
- [22] Anonymous , journal =. Context as a Tool: Context Management for Long-Horizon on Pith
- [23] Zhang, Yuhan and Guo, Zhiyuan and Zeng, Ziheng and Wang, Wei and Wu, Wentao and Xu, Lijie , journal = on Pith
- [24] Xiong, Zidi and Lin, Yuping and Xie, Wenya and He, Pengfei and Liu, Zirui and Tang, Jiliang and Lakkaraju, Himabindu and Xiang, Zhen , booktitle =. How Memory Management Impacts on Pith
- [25] Fang, Runnan and others , journal = on Pith
- [26] Anonymous , journal = on Pith
- [27] arXiv preprint arXiv:2606.24597 , year = on PithPith's reviewarXiv
- [28] Tan, Juntao and Yang, Liangwei and Zhao, Wenting and Qiu, Jielin and Zhu, Ming and Murthy, Rithesh and Savarese, Silvio and Wang, Huan and Heinecke, Shelby and Xiong, Caiming , booktitle =. A Lightweight, Domain-Adaptive Memory System for on Pith
- [29] Shaham, Uri and Segal, Elad and Ivgi, Maor and Efrat, Avia and Yoran, Ori and Haviv, Adi and Gupta, Ankit and Xiong, Wenhan and Geva, Mor and Berant, Jonathan and Levy, Omer , booktitle = on Pith
- [30] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models , author =. International Conference on Learning Representations (ICLR) , year = on Pith
- [31] Cognitive Architectures for Language Agents , author =. Transactions on Machine Learning Research (TMLR) , year = on Pith
- [32] Yang, John and Jimenez, Carlos E. and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , booktitle = on Pith
- [33] A Survey of Context Engineering for Large Language Models , author=. 2025 , eprint= on Pith
- [34] User as Code: Executable Memory for Personalized Agents , author =. arXiv preprint arXiv:2606.16707 , year = on PitharXiv
- [35] The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems , author=. arXiv preprint arXiv:2605.21997 , year= on PithPith's reviewarXiv
- [36] Toward Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability , author =. arXiv preprint arXiv:2608.06503 , year = on PithPith's reviewarXiv
- [37] Slipstream: Asynchronous Context Compaction with Trajectory-Grounded Validation , author =. arXiv preprint arXiv:2605.08580 , year = on PithPith's reviewarXiv
- [38] Hindsight-Gated Memory Writing with Local Counterfactual Utility , author =. arXiv preprint arXiv:2606.16285 , year = on PithPith's reviewarXiv
- [39] arXiv preprint arXiv:2606.30005 , year = on PithPith's reviewarXiv
- [40] arXiv preprint arXiv:2607.23809 , year = on PithPith's reviewarXiv
- [41] Memory-as-Action: Context Editing as Policy Actions for Long-Horizon Agents , author =. Findings of the Association for Computational Linguistics (ACL) , year = on Pith
- [42] Sun, Weiwei and others , journal =. Scaling Long-Horizon on Pith
- [43] Fidelity Before Structure: Verbatim Chunks Beat Lossy Artifact Extraction , author =. arXiv preprint arXiv:2601.00821 , year = on PithPith's reviewarXiv
- [44] Ji, Yichao , year =. Context Engineering for on Pith
- [45] arXiv preprint arXiv:2606.17016 , year = on PithPith's reviewarXiv
- [46] Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics , author =. arXiv preprint arXiv:2603.01209 , year = on PithPith's reviewarXiv
- [47] The Replay Gap: Why Log-Stitching Replay Scores the Wrong World , author =. Conference on Language Modeling (COLM) , year = on Pith
- [48] Sleep-time Compute: Beyond Inference Scaling at Test-time , author =. arXiv preprint arXiv:2504.13171 , year = on PithPith's reviewarXiv
- [49] Context Repositories: Version-Controlled Memory for Agents , author =. 2026 , howpublished = on Pith
- [50] Regret Bounds for Online Agent Memory Eviction , author =. arXiv preprint arXiv:2607.00394 , year = on PithPith's reviewarXiv
- [51] The Bitter Lesson of Tool Calling , author =. arXiv preprint arXiv:2608.06370 , year = on PithPith's reviewarXiv
- [52] Externalization in. arXiv preprint arXiv:2604.08224 , year = on PithPith's reviewarXiv
- [53] Zep: A Temporal Knowledge Graph Architecture for Agent Memory , author =. arXiv preprint arXiv:2501.13956 , year = on PithPith's reviewarXiv
- [54] Memento: Teaching llms to manage their own context , author=. arXiv preprint arXiv:2604.09852 , year= on PithPith's reviewarXiv
- [55] ACM International Conference on Mobile Systems, Applications, and Services (MobiSys) , year = on Pith
- [56] Self-Compacting Language Model Agents , author =. arXiv preprint arXiv:2606.23525 , year = on PithPith's reviewarXiv
- [57] Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik R. , booktitle = on Pith
- [58] Zheng, Yuxiang and Fu, Dayuan and Hu, Xiangkun and Cai, Xiaojie and Ye, Lyumanshan and Lu, Pengrui and Liu, Pengfei , journal = on Pith
- [59] BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents , author=. 2025 , eprint= on Pith
- [60] Tavakoli, Mohammad and Salemi, Alireza and Ye, Carrie and Abdalla, Mohamed and Zamani, Hamed and Mitchell, J Ross , journal =. Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in on Pith
- [61] Mem0 Memory Benchmarks ( on Pith
- [62] Chhikara, Prateek and Khant, Dev and Aryan, Saket and Singh, Taranjeet and Yadav, Deshraj , journal =. Mem0: Building Production-Ready on Pith
- [63] Bartholomew, Ben , howpublished =. Hindsight Is \#1 on on Pith
- [64] Rasmussen, Preston and Paliychuk, Pavlo and Beauvais, Travis and Ryan, Jack and Chalef, Daniel , journal = on Pith
- [65] Chhikara, Prateek and Khant, Dev and Aryan, Saket and Singh, Taranjeet and Yadav, Deshraj , journal = on Pith
- [66] Memory Evaluation , author = on Pith
- [67] Honcho: Memory Infrastructure for Stateful Agents , author = on Pith
- [68] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle = on Pith
- [69] Programmatic Tool Calling , author = on Pith
- [70] Harbor: A Framework for Building and Running Agent Evaluations at Scale , author = on Pith
- [71] Remember me, refine me: A dynamic procedural memory framework for experience-driven agent evolution , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages= on Pith
- [72] Executable Code Actions Elicit Better LLM Agents , author=. 2024 , eprint= on Pith
- [73] Lies, Damn Lies, and Statistics: Is Mem0 Really. 2025 , howpublished = on Pith
- [74] Revisiting. 2025 , howpublished = on Pith
- [75] 2026 , howpublished = on Pith
- [76] Barnes, Tyler , year =. Observational Memory: 95\ on Pith
- [77] Exabase Reports State-of-the-Art Results on. 2026 , month = jul, day = on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
How the literature uses it
What this paper does with the arXiv papers it cites
What it cites · 7 sentences without a confident label
Uses the method or tool of 1 paper
-
Scroll uses a CodeAct-style interface [ref] for both task execution and context construction.
Executable Code Actions Elicit Better LLM Agents · Programmatic Context Construction · 0.98
Compares with 1 paper
-
A comparison against the best published numbers from the LOCA paper [ref] and its leaderboard is in Appendix [ref].
2602.07962 · Comparison with Existing Systems · 0.93
Cites as background 13 papers
-
External memory systems extract selected facts or episodes into a separate store and later retrieve them through semantic or structured interfaces [ref] .
MemGPT: Towards LLMs as Operating Systems · Introduction · 0.85
-
CodeAct introduced executable Python as a general action interface for LLM agents [ref] , while programmatic tool-calling and code-execution systems allow tool results to remain in sandbox variables and enter context only through selected projections [ref] .
Executable Code Actions Elicit Better LLM Agents · Related Work · 0.92
-
Prior systems have explored verbatim recall storage, event-sourced interaction logs, lossless pointers, and provenance-linked memory [ref] .
The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems · Related Work · 0.95
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment