Structurally Aligned Subtask-Level Memory for Software Engineering Agents

Kangning Shen; Jingyuan Zhang; Chenxi Sun; Wencong Zeng; Yang Yue

cs.SEFeb 25, 2026

Structurally Aligned Subtask-Level Memory for Software Engineering Agents

Authors: Kangning Shen, Jingyuan Zhang, Chenxi Sun, Wencong Zeng, Yang Yue

Organizations: Kuaishou Technology, Beijing, China

Abstract

Large Language Models (LLMs) have demonstrated significant potential as autonomous software engineering (SWE) agents. Recent work has further explored augmenting these agents with memory mechanisms to support long-horizon reasoning. However, these approaches typically operate at a coarse instance granularity, treating the entire problem-solving episode as the atomic unit of storage and retrieval. We empirically demonstrate that instance-level memory suffers from a fundamental granularity mismatch, resulting in misguided retrieval when tasks with similar surface descriptions require distinct reasoning logic at specific stages. To address this, we propose Structurally Aligned Subtask-Level Memory, a method that aligns memory storage, retrieval, and updating with the agent's functional decomposition. Extensive experiments on SWE-bench Verified demonstrate that our method consistently outperforms both vanilla agents and strong instance-level memory baselines across diverse backbones, improving mean Pass@1 over the vanilla agent by +4.7 pp on average (e.g., +6.8 pp on Gemini 2.5 Pro). Performance gains grow with more interaction steps, showing that leveraging past experience benefits long-horizon reasoning in complex software engineering tasks.

Figures & tables

Figure 1: Comparison of Instance-Level vs. Subtask-Level Memory. Top: Instance-level memory relies on global task similarity, causing reasoning interference when tasks share surface goals but require different reasoning logic. Bottom: Our method retrieves by matching stage-consistent subtask intents, enabling precise experience reuse even across globally dissimilar tasks.

Figure 2: Overview of the Structurally Aligned Subtask-Level Memory method. Unlike instance-level approaches, our method aligns memory operations with the agent’s functional decomposition (e.g., Edit ). The process operates in two key phases: (1) Retrieval ( $R$ ): Initialized by the Subtask Intent , the agent queries the Memory State ( $S_{\text{sub}}$ ). A Category Filter followed by a Similarity Match retrieves a structurally relevant historical anchor to provide Augmented Context . (2) Update ( $U$ ): The Subtask Trajectory is processed by an Extractor to summarize abstract, transferable insights, which are stored as a New Memory Entry .

Backbone Model	Vanilla Agent		Instance-level Mem		Subtask-level Mem (Ours)		Improvement ( $\Delta$ )
Backbone Model	Best@3	Avg. Pass@1	Best@3	Avg. Pass@1	Best@3	Avg. Pass@1	Abs.	Rel.
Gemini 2.5 Flash	36.8	36.3 $\pm$ 0.76	39.8	37.8 $\pm$ 2.00	43.2	41.9 $\pm$ 1.17	+5.6	+15.4%
Gemini 2.5 Pro	54.8	53.5 $\pm$ 1.21	56.0	55.1 $\pm$ 1.50	61.2	60.3 $\pm$ 1.29	+6.8	+12.7%
Claude 3.7 Sonnet	52.8	52.2 $\pm$ 0.72	53.6	51.1 $\pm$ 2.27	57.2	56.1 $\pm$ 1.33	+3.9	+7.5%
Claude 4.0 Sonnet	64.4	63.5 $\pm$ 0.81	63.8	63.3 $\pm$ 0.81	66.8	65.8 $\pm$ 0.92	+2.3	+3.6%

Table 1: Main results on SWE-bench Verified. We report the Best@3 (the maximum Pass@1 score across three independent runs) and the Avg. Pass@1 (mean

\pm

sample std. dev.). Inst-Mem refers to the instance-level memory baseline reproduced from Ouyang et al. (2025) .

\Delta

and Rel. denote the absolute and relative improvement of our subtask-level memory method over the Vanilla baseline.

Configuration	Pass@1 (%)	$\Delta$
Vanilla Agent	52.2	-
Structured prompting only	53.2	+1.0
Ours (Full Method)	56.1	+3.9

Table 2: Ablation: With Memory vs. Structured Prompting Only.

Retrieval Strategy	Pass@1 (%)	$\Delta$
Vanilla Agent	52.2	-
No Category Filter	53.8	+1.6
Ours (Category Isolated)	56.1	+3.9

Table 3: Ablation: Category Isolation vs. Global Retrieval.

Memory Content	Pass@1 (%)	$\Delta$
Vanilla Agent	52.2	-
Raw Trajectory	53.4	+1.2
Ours (Abstract Insight)	56.1	+3.9

Table 4: Ablation: Abstract Insights vs. Raw Trajectories.

Figure 3: Temporal Dynamics of Experience Accumulation. Net gain (

\Delta

Resolved) relative to the baseline across sequential bins of 100 instances. The trend illustrates the transition from a sparse memory state (1-200) to accelerated knowledge transfer (301-500) as experience accumulates.

Figure 4: Pass@1 Improvement by Complexity. Instances are grouped by baseline trajectory length (step count). The method delivers disproportionate gains (+8.7%) on Hard tasks (

>28

steps), demonstrating its utility in mitigating long-horizon reasoning failures compared to the baseline.

Figure 5: Memory Retrieval Frequency. The distribution follows a long-tail pattern: a small “head” of generic memories is retrieved frequently, while a vast “tail” of over 400 single-use memories captures instance-specific edge cases.

Figure 6: Functional Distribution of Memories. The semantic category of retrieved memories shifts dynamically across phases. Diagnostics dominate the Analyze phase (62.0%), shifting to Implementation during Edit (62.0%) and Strategy during Verify (64.0%).

Figure 7: Case study on sympy__sympy-13757 . While the instance-level baseline fails due to coarse-grained retrieval of generic concepts, our method retrieves precise, subtask-specific insights to successfully resolve the reflected multiplication issue.

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 368 papers

Enhancing Software Engineering Through Closed-Loop Memory Optimization
Jun 4, 2026Xuehang Guo, Zora Zhiruo Wang, Qingyun Wang +2Software Engineering Agents LLM Agent Evaluation
1William & Mary · 2Carnegie Mellon University · 3OpenHands +1
SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents
Jun 26, 2026Shuzheng Gao, Wenhao Zeng, Zhaojian Yu +5Software Engineering Agents Memory Management for LLM Agents
The Chinese University of Hong Kong, China · Shanghai Jiao Tong University, China · Tsinghua University, China +1
Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills
Jun 5, 2026Chuan Xiao, Zhengbo Jiao, Shaobo Wang +5Software Engineering Agents LLM Agent Skill Learning
2Shanghai Jiao Tong University · 1AI Data, Alibaba Group

Jun 4, 2026J/K move · Enter open · S save

Enhancing Software Engineering Through Closed-Loop Memory Optimization

Xuehang Guo, Zora Zhiruo Wang, Qingyun Wang, Graham Neubig+1

1William & Mary · 2Carnegie Mellon University · 3OpenHands+1

Software Engineering Agents LLM Agent Evaluation Memory-Augmented Agents Agent Memory Management

Open paper AskPDF arXiv

Large language models (LLMs) have enabled powerful software engineering (SE) agents capable of navigating complex codebases and resolving real-world issues. However, these agents remain fundamentally episodic: they fail to retain, refine, and reuse experiences across tasks, repeatedly reconstructing context from scratch and reproducing similar mistakes. Even with memory support, they offer no remedy for the absence of a principled, task-agnostic memory utility, making them difficult to evaluate rigorously or generalize across agents and settings. To tackle these limitations, we introduce \ours, a closed-loop framework for memory augmentation in SE agents. \ours grounds memory utility in validated downstream impact, establishing utility as both a task-agnostic evaluation benchmark and an annotation-free optimization signal. Through complementary evaluation on single-episode and cross-episode memory augmentation, results demonstrate that \ours consistently improves SE agents across settings, achieving absolute gains of up to

\uparrow5.25\%

in success rate and

\uparrow4.63\%

in resolve efficiency, while substantially reducing computational cost by

\geq9.79\%

. Our project page: https://xhguo7.github.io/MemOp/.

Abstract

LLM Agent Memory

Explore similar work

Enhancing Software Engineering Through Closed-Loop Memory Optimization

SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents

Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills