StateComp: Learning When to Compress History in Long Horizon Agents

Mingxuan Wang; Hongyue Chen; Yinglong Guo; Fei Luo; Chao Ning; Bo Wang; Guorun Yao; Yanbiao Ma

cs.AISep 23, 2026

StateComp: Learning When to Compress History in Long Horizon Agents

Authors: Mingxuan Wang, Hongyue Chen, Yinglong Guo, Fei Luo, Chao Ning, Bo Wang, Guorun Yao, Yanbiao Ma,

Organizations: TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University

Abstract

Long-horizon agents continuously accumulate interaction history during task execution, yet the importance of past interactions changes as the agent state evolves. Existing context management methods largely compress history based on fixed windows, periodic schedules, or current relevance, overlooking a more fundamental question: when has a past interaction become safe to replace? Premature compression may remove information still needed for future actions, while overly conservative retention leads to substantial context overhead. To address this, we propose State Conditioned Compression (StateComp), a framework that determines when historical interactions can be safely compressed according to the current agent state. StateComp constructs KEEP and READY supervision through a two-stage annotation procedure and trains an imbalance-aware router on hidden representations from a frozen language model. A bounded state representation further reduces the cost of evaluating long histories, while adjacent READY interactions are grouped into continuous spans and replaced with compact summaries during execution. Experiments on WorkBuddyBench show that StateComp reduces total agent and summarization tokens by 52.27% while maintaining task performance, and achieves a 12.67-fold speedup in representation extraction.

Figures & tables

Figure 1: Overview of StateComp . Structured annotations and a second boundary review produce supervision. A router learns from hidden states of a frozen representation model. Online predictions are combined with span length and token gates before summaries replace raw interactions. Historical and current state features are used jointly, as described in Section 4.1 .

Figure 2: Compression dynamics across checkpoints in an example trajectory.

Method	Max. tokens	Wall clock (s)	Peak GPU (GiB)	Workspace (GiB)	Test F1
Full history	203,675	76.40	29.426	11.867	0.7000
Bounded	5,120 (-97.49%)	6.03 (12.67 $\times$ )	18.620 (-36.72%)	1.062 (-91.05%)	0.7818 (+0.08)

Table 1: Efficiency and quality of bounded state representations. Max. tokens reports the largest recorded input; runtime and memory are measured on the same seven paired states.

Figure 3: Compression opportunities over different trajectory stages.

Figure 4: Router score distributions.

Method	Code	Office	Sec.	Web	Avg.	Tokens (M)
Baseline	76.99	81.84	47.76	72.14	69.87	698.17
StateComp	77.12 $\uparrow$ 0.13	82.77 $\uparrow$ 0.93	47.31 $\downarrow$ 0.45	73.14 $\uparrow$ 1.00	70.26 $\uparrow$ 0.39	333.24

Table 2: Full WorkBuddyBench comparison on 260 tasks. Scores are mean reward multiplied by 100. Tokens include the main agent and summary calls. Colored changes show score differences relative to the baseline for all domains.

Method	Code	Office	Sec.	Web	Avg.	Tokens (M)
Baseline	72.93	81.64	44.58	71.00	67.54	75.50
Sliding Window ( $K=5$ )	66.81 $\downarrow$ 6.12	67.43 $\downarrow$ 14.21	32.32 $\downarrow$ 12.26	59.00 $\downarrow$ 12.00	56.39 $\downarrow$ 11.15	118.08
Sliding Window ( $K=10$ )	74.37 $\uparrow$ 1.44	73.05 $\downarrow$ 8.59	40.07 $\downarrow$ 4.51	61.00 $\downarrow$ 10.00	62.12 $\downarrow$ 5.42	79.39
Sliding Window ( $K=20$ )	79.31 $\uparrow$ 6.38	85.11 $\uparrow$ 3.47	40.28 $\downarrow$ 4.30	70.00 $\downarrow$ 1.00	68.67 $\uparrow$ 1.13	80.82
Periodic Summary ( $n=3$ )	65.44 $\downarrow$ 7.49	67.66 $\downarrow$ 13.98	26.32 $\downarrow$ 18.26	82.66 $\uparrow$ 11.66	60.52 $\downarrow$ 7.02	88.57
Periodic Summary ( $n=5$ )	83.60 $\uparrow$ 10.67	75.74 $\downarrow$ 5.90	40.03 $\downarrow$ 4.55	81.32 $\uparrow$ 10.32	70.17 $\uparrow$ 2.63	94.85

Table 3: Context management results on Eval40. Scores are reward multiplied by 100. Small colored arrows show changes relative to the DeepSeek baseline. Bold and underlined values indicate the best and second best results.

Backbone	Method	Code	Office	Sec.	Web	Avg.	Tokens (M)
	Base	55.30	79.24	60.24	46.00	60.19	41.00
Hunyuan	StateComp	64.31 $\uparrow$ 9.01	82.20 $\uparrow$ 2.96	60.14 $\downarrow$ 0.10	49.00 $\uparrow$ 3.00	63.91 $\uparrow$ 3.72	12.83
	Base	75.25	83.42	45.28	86.00	72.49	49.75
Qwen3.8-Flash	StateComp	93.46 $\uparrow$ 18.21	88.71 $\uparrow$ 5.29	41.20 $\downarrow$ 4.08	88.00 $\uparrow$ 2.00	77.84 $\uparrow$ 5.35	38.23
	Base	79.49	66.13	34.67	66.51	61.70	47.58
GLM-5.3-Flash	StateComp	66.15 $\downarrow$ 13.34	69.90 $\uparrow$ 3.77	38.64 $\uparrow$ 3.97	85.00 $\uparrow$ 18.49	64.92 $\uparrow$ 3.22	38.49

Table 4: Cross-model evaluation on Eval40. Scores are reward multiplied by 100. Colored arrows show changes relative to the same backbone without StateComp. Tokens report the total model usage in millions. Each domain contains ten tasks, so Avg. is the mean of the four domain scores.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$S_{i}$	Complete historical interaction
$P_{k}$	Uncompressed reference prefix before interaction $k$
$H_{k},\mathcal{I}_{k}$	Effective history and identifiers of surviving raw interactions
$y_{i,k}$	Reference label: KEEP $=0$ , READY $=1$
$t_{i}^{*}$	Earliest supported compression checkpoint
$F_{\phi},g_{\theta}$	Frozen representation model and trained router

Appendix

Table A1 : Notation for supervision and online compression.

Quantity	Value
Trajectories	300
Recorded states	7,931
Checkpoint groups	7,631
Interaction–state pairs	244,526
KEEP pairs	228,893
READY pairs	15,633

Appendix

Table A2 : Compression supervision and evaluation partitions.

Figure A1 : Variation in compression opportunities. Mean and median READY fractions and the 10th–90th percentile range across trajectories.

Figure A2 : Compression delay and span dynamics. Left: normalized delay to READY by interaction position. Right: relative span count, span length, and coverage over task progress.

Figure A3 : Prediction across training fractions and trajectory-length groups. Thin curves show individual folds and the thick curve their mean.

Figure A4 : Held-out precision and coverage. The 98% line is a reference target, not an attained guarantee.

Figure A5 : Calibration diagnostics. The displayed ECE values are 0.035, 0.038, and 0.011 for raw, temperature-scaled, and Platt-calibrated scores.

Method	Total tokens (M)	Summary tokens (M)
Baseline	75.50	0
StateComp	43.76	0.077
Sliding Window $K=5$	118.08	0
Sliding Window $K=10$	79.39	0
Sliding Window $K=20$	80.82	0
Periodic Summary $n=3$	88.57	4.50

Appendix

Table A9 : Eval40 total tokens and summary overhead. Counts are in millions. Summary tokens are included in the total; totals follow the precision of the main-text comparison.

Model	Method	Total tokens (M)	Summary tokens (M)
Hunyuan	Base	41.00	0
Hunyuan	StateComp	12.83	0.130
Qwen3.8-Flash	Base	49.75	0
Qwen3.8-Flash	StateComp	38.23	0.236
GLM-5.3-Flash	Base	47.58	0
GLM-5.3-Flash	StateComp	38.49	0.101

Appendix

Table A11 : Cross-model total tokens and summary overhead. Counts are in millions. Summary tokens are included in the total.

Context Compression

Context compression reduces the information or memory a model must carry while preserving what it needs to answer questions and act.

See all 90 papers

When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression
Sep 24, 2026Mingxuan Wang, Fei Luo, Bo Wang +6Long-Context Modeling Long-Horizon LLM Agents
TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University
Memory Control Signals Emerge Before Action in Long Horizon Agents
Sep 23, 2026Mingxuan Wang, Guorun Yao, Fei Luo +6Agent Memory Management Long-Horizon LLM Agents
TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University
Towards Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability
Aug 6, 2026Guanghui Min, Liang Wu, Mayank Darbari +2AI Agent Evaluation Long-Horizon LLM Agents
Department of Computer Science, University of Virginia, Charlottesville, VA, USA · Nokia, Sunnyvale, CA, USA

Sep 24, 2026J/K move · Enter open · S save

When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression

Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao+5

TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University

Long-Context Modeling Long-Horizon LLM Agents CoT Compression Context Compression

Open paper AskPDF arXiv

Long horizon language model agents continually accumulate reasoning history, increasing context length and inference cost even after earlier decisions have been executed and observed. Unlike static Chain of Thought compression, removing historical reasoning can change future actions and the resulting interaction trajectory. We study when such reasoning can be safely forgotten. We propose Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training free online method that ranks reasoning blocks using frozen proxy entropy while preserving actions, tool calls, and observations. On 260 WorkBuddyBench tasks, ICLR improves average reward from 0.699 to 0.718, while reducing input, output, and cache read tokens by 25.5%, 14.4%, and 33.3%, respectively. Ablations reveal trajectory amplification, where local reasoning deletion produces nonlinear changes in total computation by altering subsequent interaction. Representation probing, activation patching, and controlled trajectory analyses further suggest that historical reasoning becomes more replaceable once task relevant derived state has been reliably externalized into code, files, tool outputs, or environmental feedback. These results characterize agent reasoning as dynamic working state rather than permanent interaction history.

Abstract

Context Compression

Explore similar work

When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression

Memory Control Signals Emerge Before Action in Long Horizon Agents

Towards Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability