Research
MemCoder: Code Agent with Structured Memory Achieves 77.8% on SWE-bench Verified
MemCoder equips code agents with structured memory built from historical commit history, distilling intent-to-code mappings and applying self-refinement via verification feedback. Using DeepSeek-V3.2 as backbone, it boosts SWE-bench Verified resolved rate from 68.4% to 77.8%, a new SOTA. The key insight is that co-evolution with project history—rather than stateless code generation—unlocks the potential of general models on complex real-world software engineering tasks.
↳ Follow the thread