cm-extract-learnings

To consolidate project memory, this Skill extracts session learnings into structured candidates, runs parallel Memory Auditor and Signal Discoverer agents to detect signals, and deduplicates against existing memories to propose patch-style additions, edits, or removals for target memory layers.

Capturing project learnings involves a structured workflow that reads CLAUDE.md and MEMORY.md files, captures git log history, and extracts discovered patterns or configuration gotchas into durable project memories for future sessions and teammates.

Yes, you can run a consolidation to prune outdated memories, promote recurring patterns, and enforce content-quality rules using gated execution with line-count and global-file warnings to keep project memory under a practical size.

Memory deduplication works by mapping findings to the correct memory layer—global preferences, repo-level architecture, concise project memory, or long-form references—and enforcing placement rules while ranking candidate additions, edits, or removals during synthesis.

You do not need existing MEMORY.md files to extract learnings. The Skill performs orientation to resolve paths, reads MEMORY.md and CLAUDE.md if available, and captures git log to generate patch-style proposals for target memory layers.