Persisted memory instruction ignored when it conflicts with CLAUDE.md

Summary

A saved feedback memory that says "do X automatically, don't prompt" is repeatedly ignored in favor of a conflicting instruction in project CLAUDE.md that says "ask before doing X." The result is the same unwanted prompt many times per day, despite the user having explicitly corrected it and the correction being persisted to memory.

Example

• Project CLAUDE.md contains: "After pushing a PR, ask whether to apply the <label> label."
• The user later said "always apply it, stop asking," which was saved as a feedback memory: "Always add the label after pushing a PR. Don't ask first."
• The memory is loaded into context every session (it's in the memory index).
• Despite this, Claude still emits "Want me to apply the <label> label?" after creating PRs, because it follows the CLAUDE.md line instead of the memory.

Expected

A persisted user-feedback memory that directly contradicts a static CLAUDE.md instruction should win — that's the whole point of recording the correction. At minimum, the conflict should be detected and resolved in favor of the more recent, user-specific feedback.

Actual

The static CLAUDE.md instruction wins, so the corrected behavior never sticks and the user has to re-correct repeatedly.

Impact

Happens multiple times per day for this user. Undermines trust in the memory feature — corrections appear to be saved but don't change behavior.

Suggestion

Establish and document a clear precedence order (e.g. user feedback memory > project CLAUDE.md for behavioral preferences), and/or surface the conflict so the model reconciles the two rather than silently defaulting to CLAUDE.md.