Search papers, labs, and topics across Lattice.
MEMCoder addresses the challenge of using LLMs for code generation with private libraries by creating a multi-dimensional evolving memory of API usage guidelines learned from the model's own problem-solving experience. This memory captures both task-level API coordination patterns and API-level parameter constraints, which are often missing from static documentation. By retrieving both static documentation and relevant historical guidelines during inference, and updating the memory based on execution feedback, MEMCoder achieves a 16.31% average absolute pass@1 gain over standard RAG on private library code generation benchmarks.
LLMs can bootstrap their understanding of private APIs by autonomously learning from their own coding attempts, outperforming retrieval-augmented generation by 16% on code generation tasks.
Large Language Models (LLMs) excel at general code generation, but their performance drops sharply in enterprise settings that rely on internal private libraries absent from public pre-training corpora. While Retrieval-Augmented Generation (RAG) offers a training-free alternative by providing static API documentation, we find that such documentation typically provides only isolated definitions, leaving a fundamental knowledge gap. Specifically, LLMs struggle with a task-level lack of coordination patterns between APIs and an API-level misunderstanding of parameter constraints and boundary conditions. To address this, we propose MEMCoder, a novel framework that enables LLMs to autonomously accumulate and evolve Usage Guidelines across these two dimensions. MEMCoder introduces a Multi-dimensional Evolving Memory that captures distilled lessons from the model's own problem-solving trajectories. During inference, MEMCoder employs a dual-source retrieval mechanism to inject both static documentation and relevant historical guidelines into the context. The framework operates in an automated closed loop by using objective execution feedback to reflect on successes and failures, resolve knowledge conflicts, and dynamically update memory. Extensive evaluations on the NdonnxEval and NumbaEval benchmarks demonstrate that MEMCoder substantially enhances existing RAG systems, yielding an average absolute pass@1 gain of 16.31%. Furthermore, MEMCoder exhibits vastly superior domain-specific adaptation compared to existing memory-based continual learning methods.
Tsinghua AI