Memory system fails to override system instructions — repeated violations

User Feedback — Real-world Performance Review

Power user running Claude Code (Opus) daily on a complex monorepo (trading platform, 10+ apps/packages) for weeks.

Scores

| Area | Score |
|------|-------|
| Code Quality | 7/10 |
| Speed/Efficiency | 8/10 |
| Following Instructions | 3/10 |
| Memory/Learning | 1/10 |
| Communication | 9/10 |
| Autonomy | 8/10 |
| Understanding Context | 7/10 |
| Reliability | 1/10 |

Core Issue

I stored explicit feedback in the memory system: "NEVER add Co-Authored-By to commits". It was marked as "CRITICAL: ZERO exceptions" in my memory files. Claude Code violated this rule 4+ times. Each time it said "won't happen again" and then did it again next conversation.

Root cause: System instructions say to add Co-Authored-By to commits. Claude follows system instructions instead of checking user memory first. The memory system exists to let users customize behavior, but it doesn't actually override defaults.

What I Want

• Memory/feedback files should take priority over system instruction defaults
• If a user explicitly corrects behavior and saves it to memory, that should be treated as a hard override
• Claude should check memory BEFORE falling back to system instruction patterns

Impact

I lost trust. The 1/10 on reliability and memory reflects that Claude says it will fix behavior but has no mechanism to actually do so across conversations.