GitHub's rac-core project, branded as Lore, solves a growing problem in AI-assisted coding: agents repeatedly questioning decisions teams already made. The tool captures requirements, decisions, and designs as plain Markdown in repositories, then serves them read-only to Claude Code and similar agents via the Model Context Protocol (MCP). Unlike fuzzy search tools, it provides deterministic retrieval—returning exact decisions rather than probabilistic matches.

The engine classifies Markdown artifacts using YAML frontmatter and validates them against type-specific schemas. Validation happens in CI, rejecting malformed content or references to superseded decisions before they merge. This creates a source of truth that agents can cite rather than violate, built on the Requirements as Code (RAC) methodology.

Teams using heavy agent workflows like Claude Code or Cursor will find rac-core particularly valuable for maintaining consistency across automated coding sessions. The deterministic approach means agents reference the same decisions every time, eliminating the flakiness of embedding-based retrieval systems. Early adopters can install via pip and scaffold their first decision artifact immediately.