Technical Summary: SWE-MeM

Problem Statement

Long-horizon software engineering (SWE) agents face significant challenges when managing lengthy and noisy interaction histories under limited context budgets. Existing agents typically follow a ReAct-style append-only paradigm, where interaction history accumulates continuously. In SWE tasks, this results in trajectories that often exceed the context window of Large Language Models (LLMs) or become too noisy for effective reasoning.

Current memory management approaches suffer from two primary limitations:

1. Rigidity: Existing methods rely on static compression workflows, such as threshold-based summarization, subtask-level folding, or learned compression tools with fixed rules. These approaches fail to adapt to the heterogeneous value of information within a trajectory. For instance, lengthy execution logs might be safely compressed at any point, whereas code inspected early in a trajectory may remain essential for later reasoning. Fixed rules cannot distinguish between these scenarios.
2. Suboptimal Objectives: Most existing methods optimize strictly for context reduction rather than task-solving efficiency. While they may reduce peak context length, they often increase total token usage or interaction steps dramatically, leading to worse overall efficiency and performance.

Methodology

The authors propose SWE-MeM, a training framework designed to equip SWE agents with proactive and on-demand memory management capabilities. The approach consists of three core components:

1. Flexible Memory Management Tool

SWE-MeM introduces a tool that allows agents to decide when, what, and how to compress based on the current trajectory state.

• Mechanism: The agent invokes a compress tool with arguments including an analysis of progress, a start/end step number defining the span to compress, the summarized content, and a remaining work plan.
• Operation: The tool replaces a selected span of raw history (e.g., logs, failed attempts) with a high-level summary while preserving temporal ordering. Crucially, compressed summaries can themselves be selected for further compression in later steps, enabling progressive context reduction.

2. Curriculum Fine-Tuning with Synthesized Trajectories

To teach the model effective memory management, the authors developed a synthesis workflow that augments task-solving trajectories with memory-management actions.

• Trajectory Synthesis: A base model generates task rollouts, while a proprietary model acts as a "context monitor" to judge when compression is needed. Triggers are based on two categories:
  • Context Budget Pressure: A probabilistic scheduler that increases compression likelihood as the remaining context budget drops below 20%.
  • Proactive Memory Management: An LLM-as-a-judge mechanism detects situations where compression is beneficial before the budget is critical, such as subtask completion, low-information density (e.g., verbose logs), or focus degradation.
• Quality Filtering: A rule-based pipeline filters synthesized trajectories, masking defects like inappropriate reduction ratios, overly short ranges, or range misalignment.
• Curriculum Learning: Training occurs in two stages. The first stage teaches basic tool usage on a full set of filtered trajectories. The second stage focuses on high-quality proactive memory behaviors, using rubrics to ensure the analysis and compression are semantically consistent and actionable.

3. Memory-aware GRPO (Group Relative Policy Optimization)

To jointly optimize memory management and issue resolution, the authors propose a reinforcement learning variant called Memory-aware GRPO.

• Memory-aware Trajectory Splitting: Standard RL treats rollouts as continuous sequences. SWE-MeM checkpoints trajectories at compression steps, splitting them into sub-trajectories that start from the exact compressed prefix the agent would observe during inference. This ensures training-inference consistency.
• Step-level Credit Assignment: Standard GRPO assigns credit based on the final outcome. SWE-MeM introduces step-level masks to improve credit assignment:
  • Quality Checks: Masks inappropriate memory actions (e.g., wrong reduction ratios) in successful trajectories.
  • Late Management: Penalizes memory actions taken only after the budget has critically fallen.
  • Overflow Failure: Assigns penalties specifically to steps where the agent failed to invoke memory management despite a low remaining budget.
  • Invalid Usage: Masks steps with invalid tool calls to prevent learning undesirable patterns.

Key Contributions

1. Adaptive Memory Tool: An interface enabling agents to dynamically decide compression timing, scope, and summarization strategy based on real-time trajectory states.
2. Training Framework: A novel pipeline combining synthesized proactive memory trajectories, curriculum fine-tuning, and Memory-aware GRPO to jointly optimize memory decisions and task performance.
3. Performance and Efficiency: Empirical evidence showing that SWE-MeM achieves superior resolve rates while maintaining lower token usage and fewer interaction steps compared to existing baselines.

Experimental Results

The framework was evaluated on SWE-Bench Verified using Qwen3-4B-Instruct and Qwen3-Coder-30B-A3B models under a 32K context budget.

• Resolve Rates:
  • 4B Model: Achieved a 43.4% resolve rate (SFT + RL), outperforming all ReAct baselines and existing memory management methods in the <10B parameter range.
  • 30B Model: Achieved a 60.2% resolve rate (SFT + RL), surpassing existing memory management baselines (e.g., Context Folding, SWE-Compressor) which typically range between 53.8% and 58.8%.
• Efficiency: SWE-MeM achieved the highest resolve rate while maintaining the lowest reported token usage among memory management methods. It used fewer tokens than the corresponding base ReAct agent, whereas competing methods often increased total token usage.
• Generalization: The method demonstrated consistent improvements on SWE-Bench Multilingual and SWE-Bench Pro, suggesting the learned memory management behaviors transfer across programming languages and longer-horizon tasks.

Significance and Claims

The paper claims that SWE-MeM addresses the critical bottleneck of context management in long-horizon SWE agents by shifting from rigid, externally triggered compression to proactive, agent-driven memory management.

The authors emphasize that effective memory management does more than merely prevent context overflow; it provides a cleaner working context that helps models utilize task-relevant information more effectively. By training agents to recognize when information density is low or when a subtask is complete, SWE-MeM reduces distraction from redundant context without inducing excessive re-exploration. The results suggest that joint optimization of memory decisions and task-solving performance via Memory-aware GRPO is essential for achieving high performance in constrained context environments.