MemoCoder: Automated Function Synthesis using LLM-Supported Agents

TL;DR

MemoCoder is a multi-agent framework that enhances code generation by enabling collaborative problem solving, persistent learning from past fixes, and supervisory guidance, leading to improved performance on multiple benchmarks.

Contribution

It introduces MemoCoder, a novel multi-agent system with a Fixing Knowledge Set and Mentor Agent for iterative debugging and knowledge reuse in LLM-based code synthesis.

Findings

Outperforms zero-shot prompting and self-repair strategies.

Achieves 3.1% to 12.1% improvements in Pass@10.

Achieves 1.4% to 14.5% improvements in Pass@50.

Abstract

With the widespread adoption of Large Language Models (LLMs) such as GitHub Copilot and ChatGPT, developers increasingly rely on AI-assisted tools to support code generation. While LLMs can generate syntactically correct solutions for well-structured programming tasks, they often struggle with challenges that require iterative debugging, error handling, or adaptation to diverse problem structures. Existing approaches such as fine-tuning or self-repair strategies either require costly retraining or lack mechanisms to accumulate and reuse knowledge from previous attempts. To address these limitations, we propose MemoCoder, a multi-agent framework that enables collaborative problem solving and persistent learning from past fixes. At the core of MemoCoder is a Fixing Knowledge Set, which stores successful repairs and supports retrieval for future tasks. A central Mentor Agent supervises the repair process by identifying recurring error patterns and refining high-level fixing strategies, providing a novel supervisory role that guides the self-repair loop. We evaluate MemoCoder across three public benchmarks -- MBPP, HumanEval, and LiveCodeBench -- spanning a range of problem complexities. Experimental results show that MemoCoder consistently outperforms both zero-shot prompting and a Self-Repair strategy, with improvements ranging from 3.1% to 12.1% in Pass@10 and from 1.4% to 14.5% in Pass@50, demonstrating its effectiveness in iterative refinement and knowledge-guided code generation.