core-promote-learnings

Cross-project learnings are shared by scanning project contexts and promoting recurring entries to a global memory file. A skill identifies summaries appearing in at least three projects and appends them to ~/.claude/memory/MEMORY.md, creating a centralized repository of patterns and pitfalls.

Jaccard similarity clusters project learnings by computing token overlap on learned summaries. This mathematical measurement identifies recurring patterns across repositories by comparing tokenized text, ensuring only genuinely similar observations are grouped together for promotion to global memory.

To prevent duplicate entries when promoting patterns, the system reads existing memory at ~/.claude/memory/MEMORY.md before writing. It consults the current global memory file to verify that matching observations have not already been promoted, avoiding redundant additions while preserving per-project data.

Yes, you can aggregate cross-project learnings without altering individual project files. The promotion process scans project contexts to find shared patterns and writes exclusively to the global ~/.claude/memory/MEMORY.md, leaving all per-project data completely untouched and preserving local contexts.

Automatically scanning project contexts for recurring pitfalls and observations requires running a cross-project analysis workflow. The process scans all projects, aggregates learnings by type, and clusters similar items using similarity metrics, promoting groups that appear in three or more projects.

A centralized memory file for architectural decisions is needed when you want to share recurring patterns and pitfalls across multiple repositories. It provides a shared reference built from promoted cross-project learnings, ensuring consistent architectural decisions informed by historical observations.