Team memory workflow
NEPOPSX treats agent memory as a governed loop rather than an invisible side effect: discover locally, submit with provenance, review centrally, and pull only the approved set into generated files.
Import what you already have
Start with a dry run. The importer detects GitHub Copilot instructions and Claude project memory, shows candidates, and sends nothing until you run without --dry-run.
Capture new learnings
Use push-learnings when a person is deliberately submitting local reflector output. Use capture in session-end hooks or CI: it is non-interactive and designed to soft-fail instead of interrupting the developer workflow.
Review in the Learning Lab
Reviewers can inspect pending learnings, approve durable knowledge, or reject noise. Approval is the boundary between one agent's observation and organization-wide instruction.
- Check the evidence and affected services
- Prefer stable rules over one-off fixes
- Reclassify the learning when its proposed type is wrong
- Revert an approval if it produces an unwanted result
Pull the approved set
Preview remote state first, then apply it. This keeps the final repository change visible in normal code review.