Memory Control Signals Emerge Before Action in Long Horizon Agents

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

cs.AISep 23, 2026

Memory Control Signals Emerge Before Action in Long Horizon Agents

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

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

Abstract

Long horizon language model agents continuously accumulate interaction history, increasing computational cost while making relevant information harder to preserve and reuse. Existing context management methods mainly focus on how to compress or retrieve history, but largely leave open whether the model itself already represents the need for these memory operations before they occur. We study the hidden state immediately before each agent action and find that compression and recall needs are already encoded in the model's internal representations. These signals cannot be explained by simple context length or interaction progress, and they exhibit distinct formation patterns across model depth. We further show that most memory decision information is preserved in a compact recent context, while selectively restored historical evidence complements the long range dependencies that recent context misses. Based on these findings, we propose Preaction Memory with Evidence Retrieval (PaMER), which combines state guided compression with external evidence retrieval. PaMER+ further introduces step level evidence selection to recover only the historical information required by the current task. Experiments on WorkBuddyBench, across multiple context management baselines and model backbones, show that our framework substantially reduces context consumption while maintaining competitive task performance.

Figures & tables

Signal	Comp. $\uparrow$	Recall $\uparrow$
Length / Turn	0.751	0.552
Metadata	0.758	0.549
Hidden (Linear)	0.831	0.765
Hidden (MLP)	0.827	0.763

Table 1: Memory decision signals. Mean AUROC.

Figure 1: Memory-decision signals across layers. (a) AUROC of independently trained probes using the final input token or mean-pooled states. (b) Mean positive-minus-negative probe logit. Results are aggregated over five question-grouped splits.

Figure 2: Reading memory needs before an action. The diagram illustrates compression and recall as prediction targets. In the deployed controller, the compression heads read the final state from the task prefix and Recent2. The recall probe is used for analysis, not as an online retrieval gate.

Compact context
History	Tokens $\downarrow$	Comp. $\uparrow$	Recall $\uparrow$
Full Context	100.0	0.829	0.780
Recent2	26.0	0.776	0.761
+ Decision KD	26.0	0.781	0.753
Recovered history
Hidden Bank	16.8	0.769	0.730

Table 2: Compact context and historical evidence. Tokens are relative to Full Context.

Panel of Figure 3 — Figure 3: Online behavior and component comparisons. (a) Compression and evidence recall. (b) Component results. (c) Input context across interaction turns. The horizontal line marks 128K as an analysis reference, not a hard limit enforced by the controller.

Model	Code	Office	Sec.	Web	Avg.	Tokens $\downarrow$
DeepSeek-V4-Flash	77.0	81.8	47.8	72.1	69.9	2.69M
+ PaMER	75.3 (-1.7)	79.8 (-2.0)	52.3 (+4.5)	70.0 (-2.1)	69.4 (-0.4)	0.72M (-73.2%)
+ PaMER +	76.7 (-0.3)	80.9 (-0.9)	51.6 (+3.8)	76.3 (+4.2)	71.6 (+1.7)	0.76M (-71.7%)

Table 3: Full benchmark system performance and token usage. Scores are multiplied by

100

. Green and red indicate favorable and unfavorable changes from the base system.

Base and Ours
Method	Code	Office	Sec.	Web	Avg.	Tokens $\downarrow$
DeepSeek-V4-Flash 44	72.9	81.6	44.6	71.0	67.5	1.89M
+ PaMER	76.3 (+3.4)	79.9 (-1.7)	43.9 (-0.7)	72.0 (+1.0)	68.0 (+0.5)	0.90M (-52.4%)
+ PaMER +	71.3 (-1.6)	86.6 (+5.0)	47.5 (+2.9)	70.0 (-1.0)	68.9 (+1.3)	0.98M (-48.1%)
Simple Context Baselines
Sliding Window ( $K=5$ )	66.8 (-6.1)	67.4 (-14.2)	32.3 (-12.3)	59.0 (-12.0)	56.4 (-11.2)	2.95M (+56.3%)

Table 4: Comparison with context management methods on WorkBuddyBench Pilot40. All methods use the same fixed 40 tasks. Scores are multiplied by

100

. Parenthesized values show changes from the DeepSeek-V4-Flash base agent.

Model	Code	Office	Sec.	Web	Avg.	Tokens $\downarrow$
GPT-5.6-Luna	44.8	72.1	55.7	56.0	57.2	676K
+ PaMER	54.4 (+9.6)	57.8 (-14.3)	52.2 (-3.5)	55.0 (-1.0)	54.9 (-2.3)	81K (-88.0%)
+ PaMER +	41.4 (-3.4)	57.7 (-14.4)	55.3 (-0.4)	66.0 (+10.0)	55.1 (-2.1)	97K (-85.7%)
GLM-5.3-Flash 48	43.2	85.0	66.1	61.7	64.0	762K
+ PaMER	44.4 (+1.2)	86.5 (+1.5)	70.7 (+4.6)	66.7 (+5.0)	67.1 (+3.1)	318K (-58.3%)
+ PaMER +	39.6 (-3.6)	84.4 (-0.6)	70.7 (+4.6)	60.0 (-1.7)	63.7 (-0.3)	310K (-59.3%)

Table 5: Cross-model evaluation on WorkBuddyBench Pilot40. Scores are multiplied by

100

. All models are evaluated on the same fixed 40 tasks, with 10 tasks per domain. Parenthesized values show changes from the corresponding base agent. Bold indicates the best overall score and lowest token consumption within each backbone.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Target	AUROC $\uparrow$	AUPRC $\uparrow$	F1 $\uparrow$
Compression	0.829	0.690	0.639
Recall	0.780	0.597	0.558

Appendix

Table 6: Full Context reference for the compact-context analysis. Five-seed mean test metrics for the faithful-hint MLP; F1 uses per-seed validation-selected thresholds. Values are rounded only for display.

Signal	Compression		Recall
	AUROC $\uparrow$	AUPRC $\uparrow$	AUROC $\uparrow$	AUPRC $\uparrow$
Length + turn	$0.751\pm 0.066$	$0.533\pm 0.053$	$0.552\pm 0.074$	$0.336\pm 0.036$
Last tool type	$0.562\pm 0.026$	$0.335\pm 0.054$	$0.527\pm 0.056$	$0.315\pm 0.043$
Last action type	$0.583\pm 0.019$	$0.346\pm 0.057$	$0.535\pm 0.056$	$0.319\pm 0.039$
Length + turn + tool type	$0.758\pm 0.056$	$0.543\pm 0.054$	$0.549\pm 0.058$	$0.361\pm 0.066$
Recent2 TF-IDF	$0.616\pm 0.027$	$0.416\pm 0.072$	$0.610\pm 0.056$	$0.398\pm 0.075$

Appendix

Table 7: Memory-decision prediction controls. Diagnostic rows show mean

\pm

reported standard deviation over five question-grouped splits. The faithful-hint MLP reference is listed separately with mean metrics.

Setting	AUROC $\uparrow$	AUPRC $\uparrow$	F1 $\uparrow$
Compression
No hint, linear	0.831	0.670	0.601
No hint, MLP	0.827	0.665	0.606
Faithful hint, linear	0.836	0.697	0.624
Faithful hint, MLP	0.829	0.690	0.639
Recall

Appendix

Table 8: Additional probe-capacity and memory-hint measurements. Mean AUROC, AUPRC, and validation-threshold F1 over five splits. Each row denotes the named probe and hint configuration.

Representation	Pooling	Compression		Recall
		AUROC $\uparrow$	AUPRC $\uparrow$	AUROC $\uparrow$	AUPRC $\uparrow$
Block 4	Last	$0.757\pm 0.051$	$0.563\pm 0.063$	$0.722\pm 0.058$	$0.526\pm 0.087$
	Mean	$0.706\pm 0.027$	$0.483\pm 0.063$	$0.632\pm 0.073$	$0.435\pm 0.079$
Block 8	Last	$0.799\pm 0.044$	$0.623\pm 0.082$	$0.768\pm 0.053$	$0.594\pm 0.112$
	Mean	$0.735\pm 0.044$	$0.493\pm 0.068$	$0.690\pm 0.058$	$0.483\pm 0.093$
Block 16	Last	$0.810\pm 0.031$	$0.644\pm 0.082$	$0.780\pm 0.054$	$0.583\pm 0.108$

Appendix

Table 9: Layerwise linear diagnostics and the compact-context reference. Diagnostic rows show mean

\pm

reported standard deviation over five question-grouped splits. The bottom row uses the faithful-hint MLP reference from the compact-context analysis and is not part of a same-capacity layer comparison.

Comparison	Target / metric	Mean difference	Sign-flip $p$
Recent2 KD $-$ Recent2	C / AUROC	+0.0058	0.500
	R / AUROC	$-$ 0.0080	0.500
	C / F1	+0.0570	0.0625
	R / F1	+0.0278	0.3125
Raw Top5 $-$ Hidden Bank Recent5	C / AUROC	+0.0627	0.0625
	R / AUROC	+0.0389	0.1875

Appendix

Table 10: Paired compact-context differences across split seeds. C and R denote compression and recall; positive differences favor the first method. The listed tests concern compact-context interventions and do not reuse significance results from an earlier main-reference configuration.

Recent context
Representation	Tokens(%)	Compression		Recall		Agreement
		AUROC	F1	AUROC	F1	Comp.	Recall
Full Context	100.0	0.829	0.639	0.780	0.558	$\cdot$	$\cdot$
Current	8.8	0.604	0.451	0.717	0.532	$\cdot$	$\cdot$
Recent1	17.7	0.746	0.520	0.725	0.524	$\cdot$	$\cdot$
Recent2	26.0	0.776	0.541	0.761	0.520	$\cdot$	$\cdot$

Appendix

Table 11: Compact-context and historical-evidence comparison. C and R denote compression and recall; dots denote unreported teacher agreement.

Figure 4: Decision information at different input budgets. Left: recent context. Right: historical evidence recovery. Token fractions refer to the corresponding offline comparison.

Retriever	Tokens(%)	AUROC		Agreement
		Comp.	Recall	Comp.	Recall
Random Top3	47.2	0.793	0.752	0.786	0.849
Recency Top3	46.2	0.792	0.753	0.818	0.857
Selective Raw Top1	35.9	0.799	0.761	0.826	0.817
Selective Raw Top3	52.3	0.809	0.766	0.826	0.834
Selective Raw Top5	66.2	0.832	0.769	0.859	0.865

Appendix

Table 12: Offline retrieval controls. Token percentages are relative to Full Context; agreement is measured against the corresponding teacher.

Variant	Recalled tokens			Agreement
	Mean	P95	Max	Comp.	Recall
Top3 cap 8K	5,540	8,021	8,192	0.818	0.835
Top3 cap 16K	8,765	14,801	16,383	0.843	0.837
Top3 uncapped	9,167	16,509	32,245	0.838	0.838

Appendix

Table 13: Qwen Top3 evidence budgets. Mean, P95, and maximum describe recalled evidence tokens, not total prompt length.

Parameter	Value
Compression controller
Retained complete blocks	2
Feature extractor	Frozen, provenance-matched Qwen3.5-9B
Feature dimension	4,096
Compression heads	5 Recent2 decision-distillation MLPs
Vote rule	3 of 5

Appendix

Table 14: Controller settings. Character budgets, token budgets, and embedding dimensions are distinct quantities.

Condition	Required behavior
Missing session identifier	Return the original request and record the condition.
Unsafe tool protocol	Return the original request without applying a memory transformation.
Candidate below 1,024 tokens	This is not a failure: skip compression; recall from existing memory remains possible.
Probe failure or OOM	Skip compression and record the error; existing-memory recall may still be attempted.
Summary, embedding, or storage failure	Return the original request, commit no incomplete memory, and remove no raw block.
Recall transport or shape failure	Return the original request rather than a partially rewritten context.

Appendix

Table 15: Memory-operation fallback rules.

Backbone	Variant	Code	Office	Sec.	Web	Avg.	$\boldsymbol{\Delta}$
GPT-5.6-Luna	Base agent	44.8	72.1	55.7	56.0	57.2	0.0
	PaMER	54.4	57.8	52.2	55.0	54.9	-2.3
	PaMER +	41.4	57.7	55.3	66.0	55.1	-2.1
GLM-5.3-Flash	Base agent	43.2	85.0	66.1	61.7	64.0	0.0
	PaMER	44.4	86.5	70.7	66.7	67.1	+3.1
	PaMER +	39.6	84.4	70.7	60.0	63.7	-0.3

Appendix

Table 16: Cross-model results on the fixed Pilot40 task set. Scores are multiplied by 100. All variants use the same 40 tasks, with ten tasks per domain. Avg. is the equal-weight average of the four domain scores;

\Delta

is the change from the corresponding base agent.

Backbone	Variant	Tokens/task (K) $\downarrow$	Change (%)
GPT-5.6-Luna	Base agent	676	0.0
	PaMER	81	-88.0
	PaMER +	97	-85.7
GLM-5.3-Flash	Base agent	762	0.0
	PaMER	318	-58.3
	PaMER +	310	-59.3

Appendix

Table 17: Cross-model token usage on fixed Pilot40. K denotes thousands of tokens. Totals include the counted main-agent and auxiliary memory calls. Changes are relative to the corresponding base agent.

Variant	Reward $\uparrow$	Tokens/task $\downarrow$	Reported peak $\downarrow$
Base agent	0.6868	3.066M	280.1K
PaMER	0.7000	0.899M	97.5K
w/o Recall	0.6370	0.924M	107.1K
w/o Compression	0.6887	3.180M	253.3K

Appendix

Table 18: Component ablation in a separate execution batch. Reward uses the original scale.

Turn	Recorded operation	Evidence in the saved trace
19	manage_context	41 messages are summarized into memory 1.
28	manage_context	19 messages are summarized into memory 2.
36	query_memory	The agent requests memory 1 during continued search.
47	manage_context	39 messages are summarized into memory 3.
71	manage_context	49 messages are summarized into memory 4.
87	get_document	Document 20026 is retrieved as candidate supporting evidence.

Appendix

Table 19: Selected events in the illustrative trajectory.

Domain	Index	Task identifier
Code	1	api_contract-hard-markup_errors
	2	api_contract-hard-openapi_params
	3	api_contract-hard-token_errors
	4	api_contract-hard-validation_errors
	5	bug_fix-easy-a_crash_in_local
	6	bug_fix-easy-filtered_relation_queryset_arg

Appendix

Table 20: Fixed Pilot40 task identifiers.

Agent Memory Management

Agent memory management studies how AI agents store, update, retrieve, compress, and forget information across long-running tasks.

See all 446 papers

StateComp: Learning When to Compress History in Long Horizon Agents
Sep 23, 2026Mingxuan Wang, Hongyue Chen, Yinglong Guo +6Context Compression Long-Horizon LLM Agents
TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
Jul 9, 2026Yifan Wu, Lizhu Zhang, Yuhang Zhou +5LLM Agent Memory Long-Horizon Agent Tasks
Meta AI
Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory
Mar 4, 2026Zhenting Wang, Huancheng Chen, Jiayun Wang +1LLM Agent Memory Memory Management for LLM Agents
Center for Advanced AI, Accenture

Sep 23, 2026J/K move · Enter open · S save

StateComp: Learning When to Compress History in Long Horizon Agents

Mingxuan Wang, Hongyue Chen, Yinglong Guo, Fei Luo+5

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

Context Compression Long-Horizon LLM Agents Agent Memory

Open paper AskPDF arXiv

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.

Abstract

Agent Memory Management

Explore similar work

StateComp: Learning When to Compress History in Long Horizon Agents

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory