EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs

Guanhua Zhu; Bo Shen; Lei Zhang; Fadi Deek; Zhi Wei; Mengnan Du; Chenxi Yuan

cs.AIOct 14, 2026

EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs

Authors: Guanhua Zhu, Bo Shen, Lei Zhang, Fadi Deek, Zhi Wei, Mengnan Du, Chenxi Yuan

Organizations: Department of Informatics, New Jersey Institute of Technology · Department of Mechanical and Industrial Engineering, New Jersey Institute of Technology · Department of Computer Science, New Jersey Institute of Technology · School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen

Abstract

Different tasks often require interactive agents to perform similar operations at intermediate steps. Retrieving an entire past workflow can introduce irrelevant steps, whereas retrieving isolated operations can discard the dependencies needed to reuse them. We introduce EvoWeave, a framework for experience memory that represents abstract states as nodes and reusable operations as directed edges with input and output contracts, provenance, and capability references. A planner decomposes the current task into a dependency graph; operation retrieval and bounded directed search then assemble relevant experience fragments. An execution agent can inspect these fragments and their referenced capabilities, while reflection separates concrete execution evidence from reusable abstractions. We study the framework across four benchmarks, five evaluation settings, and three language model backbones. Across backbones, AppWorld Normal and Challenge task goal completion average 51.91 and 27.99, respectively, exceeding the baseline without memory by 6.99 and 4.99 percentage points. An AppWorld ablation associates directed connections with an additional improvement of 2.08 points over independent operation cards. The results motivate composition of reusable operations while exposing tradeoffs in retrieval overhead and conditions for valid reuse.

Figures & tables

Figure 1: Motivation for reusing individual operations and capability fragments. The illustrative comparison contrasts retrieving an entire historical workflow or capability file with selecting relevant operations and functions for a new task. Selective reuse aims to reduce irrelevant context while supporting composition across partially overlapping experiences. The figure is conceptual; its step counts and anticipated benefits are not experimental measurements or claims about every prior method.

Figure 2: Overview of EvoWeave. A planner generates a task DAG that guides retrieval from an experience graph, whose nodes represent states and edges represent operations. The main agent uses retrieved experience and relevant capabilities, while the reflector separates concrete step summaries from abstract operation summaries. Shared memory updates follow the declared construction or evaluation protocol. Capability and role refinement in the diagram describe the broader architecture; evaluated implementation boundaries are detailed in Appendix B .

DeepSeek-V3.2
	AppWorld Normal		AppWorld Challenge		WebArena	ALFWorld	ScienceWorld
Method	TGC	SGC	TGC	SGC	SR	Score	Score
No-Memory	$47.02_{\pm 2.75}$	$30.36_{\pm 3.30}$	$23.02_{\pm 1.97}$	$9.35_{\pm 1.83}$	$42.00_{\pm 1.36}$	$69.40_{\pm 2.41}$	$64.70_{\pm 2.52}$
Flat-Memory	$49.05_{\pm 3.03}$	$32.14_{\pm 3.68}$	$23.02_{\pm 2.16}$	$9.35_{\pm 2.01}$	$43.00_{\pm 1.49}$	$71.94_{\pm 2.59}$	$67.20_{\pm 2.77}$
AWM	$52.02_{\pm 2.92}$	$35.00_{\pm 3.62}$	$24.99_{\pm 2.12}$	$10.50_{\pm 1.99}$	$46.01_{\pm 1.44}$	$73.88_{\pm 2.47}$	$68.70_{\pm 2.67}$
ReasoningBank	$52.01_{\pm 3.99}$	$35.00_{\pm 4.95}$	$26.00_{\pm 2.92}$	$11.88_{\pm 2.76}$	$43.00_{\pm 1.97}$	$71.34_{\pm 3.46}$	$67.70_{\pm 3.66}$

Table 1: Main results reported as mean

\pm

standard deviation (SD), with the

\pm

SD terms shown as subscripts. All entries retain the original score scale from 0 to 100. Within each backbone block, the highest mean in a column is shaded and the second highest is underlined ; tied values receive the same mark.

Figure 3: Category performance averaged equally over three backbones. Each spoke uses a local linear range determined by all eight methods; brackets give its inner and outer bounds in original score units. Compare methods on the same spoke: distances across spokes, panels, and polygon areas are not directly comparable. Labels and the scaling rule are defined in Appendix C . These category aggregates are distinct from Table 1 ; see Appendix A .

Configuration	Normal TGC	Challenge TGC	Overall TGC
A: No memory	44.92	23.01	29.30
B: Whole workflow RAG	46.59	23.02	29.79
C: Operation cards	47.93	24.72	31.39
D: C + directed connections	50.74	26.51	33.47
E: C + active retrieval	49.90	25.94	32.82
F: EvoWeave (full)	51.91	27.99	34.86

Table 2: AppWorld ablation, averaged equally over backbones. Overall TGC uses weights based on task counts of 168 Normal and 417 Challenge tasks within each backbone. These configurations are specific to AppWorld; only point estimates are shown. Configurations are ordered so that the full system appears last, and the highest value in each column is shown in bold.

Experience layer	States	Operations
Mean count	27.0	157.6
Capability layer	Code	Skills	Roles
Mean nodes	185.2	82.2	6.8
Mean links to operations	157.6	362.0	99.4

Table 3: Mean memory size over the five settings. Counts describe the reported graph snapshots, not an inferred average topology. Capability links connect the corresponding node type to operations. Counts of role nodes do not establish that learning of role prompts was evaluated.

Figure 4: Mean operation overlap and reported memory referencing across task types. Each type receives equal weight; overlap comparisons are restricted to tasks of the same type. Labels show type counts, and bars show percentages on a scale from 0 to 100. Benchmark specific overlap proxies and reference rates are distinct measures (Appendix G ).

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Evaluation scope and reporting convention
Benchmarks	AppWorld Test-Normal and Test-Challenge, WebArena, ALFWorld, and ScienceWorld. These give five settings from four benchmarks.
Backbones	DeepSeek-V3.2, GPT-4.1-mini, and Qwen3.8-27B.
Deployment	Qwen served locally on an institutional HPC cluster; DeepSeek-V3.2 and GPT-4.1-mini accessed through API calls.
Methods	No-Memory, Flat-Memory, AWM, ReasoningBank, Mem0, ExpeL, A-MEM, and EvoWeave.
Main metrics	AppWorld TGC and SGC, WebArena SR, and separate scores for ALFWorld and ScienceWorld, all on scales from 0 to 100.
Memory construction	Online over the evaluation stream. Every run starts from an empty memory, uses the released benchmark task order, and is independent across the 15 backbone and setting combinations. No training or development data is used, and episode level success is judged by a language model module shared by all methods rather than by the benchmark scorer.

Appendix

Table 4: Evaluation scope and reporting conventions.

Reference parameter	Default value
Operation retrieval cosine threshold / maximum candidates	0.72 / 3
Cosine threshold for merge candidates / maximum candidates	0.82 / 3
BFS maximum depth / expansions	8 / 1,000
Maximum visible operation cards	18
Skill retrieval cosine threshold / maximum cards	0.35 / 8
Code retrieval cosine threshold / maximum cards	0.35 / 4

Appendix

Table 5: Default parameters of the reference implementation. These specify the retrieval procedure and are distinct from the settings of individual experimental runs.

Figure 5: DeepSeek-V3.2 category performance across five settings. Each spoke uses the local range printed beside its category. All eight methods are retained.

Figure 6: GPT-4.1-mini category performance across the five settings. These curves are inputs from individual backbones to the equal averages over backbones in Figure 3 .

Figure 7: Qwen3.8-27B category performance across the five settings. Each spoke uses the local score range printed beside its category.

Configuration	Normal TGC	Challenge TGC	Overall TGC
A: No memory	47.02	23.02	29.91
B: Whole workflow RAG	49.05	23.02	30.50
C: Operation cards	49.09	24.36	31.46
D: C + directed connections	52.75	25.21	33.12
E: C + active retrieval	52.20	25.67	33.29
F: EvoWeave (full)	54.05	27.00	34.77

Appendix

Table 6: DeepSeek-V3.2 AppWorld ablation.

Configuration	Normal TGC	Challenge TGC	Overall TGC
A: No memory	35.71	16.98	22.36
B: Whole workflow RAG	36.67	16.02	21.95
C: Operation cards	38.50	17.81	23.75
D: C + directed connections	39.84	19.80	25.56
E: C + active retrieval	39.11	18.86	24.68
F: EvoWeave (full)	40.60	21.01	26.64

Appendix

Table 7: GPT-4.1-mini AppWorld ablation.

Configuration	Normal TGC	Challenge TGC	Overall TGC
A: No memory	52.02	29.02	35.63
B: Whole workflow RAG	54.05	30.02	36.92
C: Operation cards	56.20	31.99	38.94
D: C + directed connections	59.63	34.51	41.72
E: C + active retrieval	58.40	33.30	40.51
F: EvoWeave (full)	61.07	35.97	43.18

Appendix

Table 8: Qwen3.8-27B AppWorld ablation.

	States	Ops.	Code		Skills		Roles
Setting			Nodes	Links	Nodes	Links	Nodes	Links
AppWorld Normal	37	173	405	173	56	142	9	117
AppWorld Challenge	51	431	467	431	139	825	9	216
WebArena	18	67	17	67	180	562	5	61
ALFWorld	9	14	11	14	6	27	1	8
ScienceWorld	20	103	26	103	30	254	10	95

Appendix

Table 9: Reported graph statistics for each setting. Ops. denotes directed operation edges. Links denotes references from capabilities to operations, not state transition edges.

Figure 8: Graph composition for the five evaluation settings. Each panel shows the experience layer and tool layer, each normalized to 100%. Labels report counts and shares within the corresponding layer; totals appear beside the bars. Solid fills indicate nodes and hatched fills indicate operation edges or capability links. The keys below each panel give all six tool components, including those with very small shares. Source counts are retained in Table 9 .

Method	Injected	Input	KV reuse (%)	Output	Rounds	s/round
No-Memory	0.00	103,690.00	70	4,925.00	18.52	9.52
Flat-Memory	26,300.80	127,710.80	78	4,829.20	17.96	9.79
AWM	19,890.00	114,140.00	78	4,479.80	16.78	9.66
ReasoningBank	11,872.00	103,192.00	77	4,341.00	16.32	9.54
Mem0	8,500.00	113,140.00	76	4,974.00	18.66	9.51
ExpeL	28,011.60	123,961.60	78	4,564.20	17.02	9.82

Appendix

Table 10: Efficiency metrics. Injected, Input, and Output denote token counts per question. Rounds and seconds per round are averages. The summaries do not fully specify the model calls included in each token metric.

Figure 9: AppWorld Normal: category overlap and reported memory referencing. Categories reflect proxies based on database models; their task counts need not partition the overall task set.

Figure 10: AppWorld Challenge: category percentages, with task counts shown alongside labels. The nine types receive equal weight in Figure 4 . Task membership can overlap across categories.

Figure 11: WebArena: overlap of evaluation operators and reported memory referencing. The five single site categories cover 764 of the 812 tasks; the remaining 48 span more than one site and are included in the main table but not in any category, as explained in Appendix A .

Figure 12: ALFWorld: overlap of high level action sets and reported memory referencing. Repeated action vocabulary does not itself establish matching objects, preconditions, or successful execution.

Figure 13: ScienceWorld: overlap of gold verb sets and reported memory referencing across 90 tasks. S1 to S10 follow the topic key in Appendix C .

Agent Memory

Agent memory lets AI systems retain, update, and use facts, experiences, and skills across sessions and long-horizon tasks.

See all 190 papers

ContextWeave: A Real-World Workflow Benchmark
Aug 5, 2026Bo Wang, Yuqian Yao, Enxi Wang +25Agent Memory Agentic Workflows
1Fudan University · 2Shanghai Innovation Institute · 3ByteDance
REM: What to Store, When to Recall, What to Forget in Experience Memory for Coding Agents
Oct 19, 2026Xinyi Lisa Chen, Pedro Herrero-Vidal, Xiaohan Chen +6Memory Management for LLM Agents Agent Memory
Yale University · Amazon
AgentEvolver: System-Wide Self-Evolution Through Task Execution
Oct 8, 2026Wentao Zhang, Fuchao Yang, Yilei Zhao +2LLM Agent Self-Improvement Self-Evolving Agents

Aug 5, 2026J/K move · Enter open · S save

ContextWeave: A Real-World Workflow Benchmark

Bo Wang, Yuqian Yao, Enxi Wang, Luozhijie Jin+24

1Fudan University · 2Shanghai Innovation Institute · 3ByteDance

Agent Memory Agentic Workflows Memory Agent Benchmarks

Open paper AskPDF arXiv

Memory is essential as language agents move from isolated tasks to long-horizon, stateful workflows, yet existing evaluations often reduce it to retrieval or question answering. We introduce ContextWeave, a longitudinal benchmark that evaluates whether recalled experience improves downstream agent performance in realistic office-work streams. ContextWeave reconstructs privacy-preserved, multi-month workflows of 14 participants into 1,005 executable tasks, including 568 core evaluation tasks, with instructions, containerized environments, trajectories, and task-specific rubrics. It measures workspace quality and alignment with participant-specific preferences, complemented by diagnostics of relevance, continuity, solvability, and robustness to misleading recall. Across six memory components under a fixed model, the strongest configuration raises Workspace Score from 68.08 to 78.20 and Preference Score from 41.50 to 70.60. With a fixed memory component, recall improves both outcomes for all five tested base models, although gains vary substantially. Our analysis shows that actionable, experience-rich memory supports workflow continuation and reduces redundant exploration more effectively than compact summaries, while it can also be more susceptible to misleading recall. These findings motivate memory systems that optimize not only retrieval relevance but also reliable use during execution.

Abstract

Agent Memory

Explore similar work

ContextWeave: A Real-World Workflow Benchmark

REM: What to Store, When to Recall, What to Forget in Experience Memory for Coding Agents

AgentEvolver: System-Wide Self-Evolution Through Task Execution