learning-capture

To save recurring preferences in CLAUDE.md, the learning-capture skill records durable corrections directly in the appropriate file. It applies a generalization gate to ensure the preference applies to future work before writing it to project or user memory.

The best way to persist workflow knowledge across sessions is using learning-capture, which routes reusable guidance to the correct memory home. It distinguishes standing instructions from reusable skills and records them where future sessions can access them.

The generalization gate checks whether a correction applies to future work or is only task-specific. It prevents one-off corrections from polluting long-term memory by only capturing generalizable corrections and recurring preferences.

Yes, learning-capture routes durable learnings to the correct memory home across project, user, plugin, and agent scopes. It decides whether to write to project memory, user memory, a plugin skill, or Harness TODO notes based on the learning type.

You should not use learning capture for task-specific corrections that do not generalize to future work. The skill enforces a generalization gate to filter out one-off instructions, preventing them from polluting long-term memory files.

Yes, learning-capture performs duplication checks before recording edits in CLAUDE.md, skill files, or Harness TODO notes. This ensures durable learnings and standing rules are not written multiple times across memory homes.