Memory compounds. Memory also contradicts itself. Two sessions ago you said use pgvector; last week you said use Pinecone; now Claude pulls the wrong one. XFCD reads the memory tree, identifies contradictions, and hands you a resolution UX that lets you approve, override, or skip. What you approve becomes the current truth — what you override gets demoted.
Part of Memory Layer
Connect with GitHub — it's freeEvery stored fact is evaluated against prior facts in the same semantic neighborhood. Direct contradictions (‘yes X / no X’), preference flips (‘use A / use B’), and architectural drift (‘we decided Y / we're replacing Y’) all surface as contradiction candidates.
XFCD never auto-resolves on first detection. You get a recommendation per contradiction with evidence from both sides, and choose approve / override / skip. Skip keeps both active. Override demotes one. Approve establishes the new truth.
Patterns of approval train XFPA. Once a rule's confidence exceeds threshold, future contradictions of that class auto-apply without HIL — while the uncertain cases still escalate. The guardrails fade where they've earned trust, never where they haven't.
You run /xfcd or XFCD fires automatically on memory writes.
Candidates surface with recommendation + both-sides evidence.
Approve / override / skip per candidate.
Approved resolutions update the tree; overrides demote the weaker entry.
XFPA learns your pattern and graduates auto-apply as confidence builds.