Auto memory should prompt before saving, not save silently

Summary

Auto memory writes to ~/.claude/projects/<project>/memory/ without user confirmation, creating hidden state that influences future sessions. This should require an explicit opt-in prompt per save, similar to how CLAUDE.md edits require permission.

Problem

• Auto memory is enabled by default and writes silently (no approval prompt)
• A previous session's corrections or one-off context can subtly bias future sessions
• Users may not realize why Claude is behaving a certain way, since the influencing state is hidden
• Stale or incorrect memories accumulate without the user's awareness
• This breaks the mental model of "fresh session, clean slate"

Proposal

Before saving an auto memory entry, prompt the user:

"I noticed [X]. Want me to save this for future sessions?"

This preserves the convenience of auto memory while giving users transparency and control — the same model already used for CLAUDE.md file edits.

Context

• CLAUDE.md edits already go through the permission system — auto memory is the only persistence mechanism that bypasses it
• The current workaround is "autoMemoryEnabled": false, but this is all-or-nothing
• Claude Code is a developer tool; its users are well-equipped to make per-save decisions