Learning to Commit: An Online Repository Memory Framework for AI to Submit Code Like Humans

This article introduces the Learning to Commit framework, whose core is the Online Repository Memory mechanism. By enabling AI to learn coding styles, internal API usage patterns, and architectural constraints from a project's historical commits, this framework addresses the problem of AI-generated code being "incompatible" with real-world projects and generates Pull Requests that better align with project practices.

Large language models excel at code generation, but their Pull Requests are often rejected when submitted to real open-source projects. The problem is not functional correctness but a lack of "organicity"—ignoring project-specific conventions, reinventing the wheel, and violating implicit architectural constraints. Existing methods rely on code snapshots, but snapshots only show the final state, lacking historical information on "why it was designed this way" and failing to capture implicit conventions like "database operations must go through the DatabaseManager class".

The core of the Learning to Commit framework is the Online Repository Memory, which consists of two phases: Skill Building Phase: Through supervised contrastive reflection, the AI generates solutions for historical Issues, compares them with human-submitted diffs, extracts skills such as coding styles, internal API usage, and architectural constraints, and accumulates them into a knowledge base. Conditional Generation Phase: When a new PR comes in, the AI retrieves relevant patterns from the skill library and generates code that aligns with the project's evolutionary history, ensuring organicity.

The evaluation uses strict time splitting: the skill library is built using commits before January 1, 2024, for the training period, and new PRs after that (ensuring they are unseen) are used for the testing period. Multi-dimensional metrics include functional correctness, code style consistency, internal API reuse rate, and rationality of modified areas. Experimental results show: increased internal API reuse rate, improved linter pass rate, enhanced architectural compliance, without sacrificing functional correctness.

The effectiveness of contrastive reflection comes from three points:

1. Active Learning: The AI actively attempts to generate and compare, which leads to deeper memory than passively reading historical commits;
2. Gap as Signal: The gap between AI-generated content and real commits reveals implicit project conventions;
3. Continuous Accumulation: The skill library grows with the increase of historical commits, and the AI's understanding of the project gradually deepens.

Typical Application Scenarios: Open-source project contributions (helping new contributors), enterprise internal development (complying with internal norms), legacy project maintenance (extracting historical knowledge). Limitations: Relies on sufficient historical data (new projects benefit less), high computational cost, possible conflicts between old and new skills during project refactoring requiring versioned management.

The Learning to Commit framework provides three implications for AI programming tools:

1. Shift from general-purpose to specialized, focusing on specific codebases;
2. Value the knowledge in code commit history;
3. Adopt reflective learning, allowing AI to actively try and learn from mistakes.

The Learning to Commit framework solves the "incompatibility" problem of AI coding agents through the online repository memory mechanism, enabling AI to not only "write correctly" but also "write like" project members. Its significance lies in revealing that true intelligence comes not only from large-scale pre-training but also from continuous learning and adaptation to specific environments. Organicity will become a key dimension for measuring the value of future AI programming tools.