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

2026-06-30 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

SWE-MeM is a novel training framework designed to equip long-horizon software engineering agents with adaptive memory management capabilities. It addresses the challenge of limited context budgets and noisy interaction histories by introducing a flexible memory tool. This tool allows agents to proactively and on-demand decide when, what, and how to compress historical context based on trajectory state, task progress, and remaining context. The framework integrates memory-management trajectory synthesis, a two-stage curriculum fine-tuning process, and Memory-aware GRPO, which jointly optimizes memory decisions and task-solving performance. Evaluated on SWE-Bench Verified, SWE-MeM achieved resolve rates of 43.4% with Qwen3-4B-Instruct and 60.2% with Qwen3-Coder-30B-A3B under a 32K context budget. It consistently outperformed existing memory management baselines in both performance and token efficiency, demonstrating its ability to provide a cleaner working context and improve task-relevant information utilization.

Key takeaway

For machine learning engineers developing long-horizon software engineering agents, consider integrating adaptive memory management like SWE-MeM. Your agents can significantly improve issue resolution rates and token efficiency by learning to proactively compress irrelevant context. Implement a flexible memory tool and train with memory-aware reinforcement learning to enable agents to make intelligent, on-demand compression decisions, rather than relying on static rules. This approach not only prevents context overflow but also provides a cleaner, more focused working context for complex tasks.

Key insights

Long-horizon coding agents can achieve superior performance and efficiency through adaptive, proactive memory management.

Principles

• Adaptive memory management improves agent efficiency.
• Jointly optimize memory decisions and task performance.

Method

SWE-MeM trains agents using synthesized memory-management trajectories, curriculum fine-tuning, and Memory-aware GRPO, which employs trajectory splitting and step-level credit assignment for joint optimization.

In practice

• Implement flexible context compression tools.
• Synthesize memory-aware training trajectories.

Topics

• Software Engineering Agents
• Memory Management
• LLM Context Management
• Reinforcement Learning
• Curriculum Learning
• SWE-Bench Benchmark

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Context Engineering for Coding Agents — Effective AI agent performance hinges on optimized context management, particularly through dynamic, wiki-based memory systems.
• MemRefine: LLM-Guided Compression for Long-Term Agent Memory — MemRefine uses an LLM judge to compress agent memory, preserving factual content within fixed storage budgets.
• Enhancing Software Engineering Through Closed-Loop Memory Optimization — MemOp grounds memory utility in validated downstream impact to optimize LLM-based SE agents.
• Multi-Agent Transactive Memory — MATM enables LLM agents to collectively store and retrieve interaction trajectories, improving task performance and efficiency across populations.
• rohitg00 / agentmemory — Agentmemory provides persistent, searchable memory for AI coding agents, enhancing efficiency and reducing redundant explanations.
• Mem-$π$: Adaptive Memory through Learning When and What to Generate — Mem-$π$ adaptively generates context-specific guidance for LLM agents on demand, outperforming static retrieval methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.