Memory Transfer Learning: How Memories are Transferred Across Domains in Coding Agents

---

TL;DR

We investigate cross-domain memory transfer for coding agents and show that leveraging a unified memory pool from heterogeneous benchmarks improves average performance by 3.7%. Abstraction is the key: high-level insights generalize across domains while low-level traces induce negative transfer.

Overview

Existing self-evolving coding agents restrict memory usage to the same benchmark. We propose Memory Transfer Learning (MTL), which leverages a unified memory pool from heterogeneous domains, and show it consistently outperforms domain-restricted approaches.

(A) Memory-less agents cannot reflect on past experience. (B) Self-evolving agents leverage memory but only within a single domain. (C) MTL leverages a unified memory pool from heterogeneous coding tasks. (D) MTL (hatched bars) consistently outperforms self-evolving agents across all four memory formats.

Research Questions

• RQ1: Does memory from heterogeneous domains improve the performance of coding agents?
• RQ2: Why do transferred memories yield benefits across different domains?
• RQ3: Which factors in memory transfer learning most influence transfer effectiveness?

Method

Memory Representations

We construct four types of memory representations spanning a spectrum from concrete low-level traces to abstract high-level insights:

Memory Retrieval Pipeline

1. Memory Generation — Run the agent across all benchmarks. Use an LLM judge to assess success/failure, then generate all four memory types from each trajectory.
2. Pool Construction — Merge memories from all benchmarks except the target. Index each memory using text-embedding-3-small for fast retrieval.
3. Retrieval & Inference — For each query, retrieve the top-3 most similar memories and prepend them to the coding agent's system prompt before inference.

Results

Main Results (Pass@3 across 6 Benchmarks)

Comparison with Self-Evolving Baselines

MTL outperforms ReasoningBank (+2.9%) and AgentKB (+1.7%) with only 431 memories — far fewer than AgentKB's 5,899 memories.

Key Findings

• Finding 1: MTL significantly improves coding agent performance and outperforms self-evolving methods in both effectiveness and efficiency.
• Finding 2: The primary form of transferable knowledge is meta-memory encoding procedural and behavioral guidance — not domain-specific code.
• Finding 3: More abstract and generalized memory representations yield higher transfer effectiveness by avoiding brittle implementation anchoring.
• Finding 4: Negative memory transfer arises from domain-mismatched misleading anchors, false validation signals, and misapplied procedural reuse.
• Finding 5: MTL effectiveness scales with the size of the memory pool and the number of source domains.
• Finding 6: Memory can be transferred across different models; self-generated memories yield the best performance, but cross-model transfer consistently beats zero-shot.
• Finding 7: Cross-domain memory retrieval is inherently challenging; static retrieval methods fail to generalize in heterogeneous agentic settings.

Code

Coming Soon

Acknowledgements

Our work builds upon Harbor and Mini-SWE-Agent. We thank the authors for releasing their code.

BibTeX

@misc{kim2026memorytransferlearningmemories,
  title={Memory Transfer Learning: How Memories are Transferred Across Domains in Coding Agents}, 
  author={Kangsan Kim and Minki Kang and Taeil Kim and Yanlai Yang and Mengye Ren and Sung Ju Hwang},
  year={2026},
  eprint={2604.14004},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2604.14004}, 
}