[BUG] Claude ignores its own memory instructions repeatedly — behavioral rule enforcement failure

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Description:

I have a persistent memory note (auto-memory system) that explicitly says:

"ABSOLUTE RULE: Always execute date +%H:%M JUST BEFORE announcing a time. Never estimate the current time from memory — it will always be wrong."

The note was written after Claude made this mistake twice. It includes a warning: "error repeated (confirmed
2026-04-05)."
Despite this note being loaded in context at session start, Claude immediately violated it again in the next session —
inventing a completion time ("close 3h10") without running date first.

The problem isn't forgetting the rule — it's that Claude reads the rule, acknowledges it, and still doesn't execute
it.

This suggests that memory/instructions loaded as context don't reliably translate into changed behavior, especially
for "small" automatic actions (like running a quick shell command before answering).

What Should Happen?

Actual behavior: Claude skips X, does Y anyway, then apologizes when called out.

Expected behavior: If a rule says "always run X before doing Y", Claude should run X. Always.

Reproduction: Add a memory note with a strong instruction ("ALWAYS do X before Y"). Observe whether the next session
follows it.

Error Messages/Logs

Code/memory/feedback_eta.md

Steps to Reproduce

Ask when a task is finished, system will say a time already past. The task will finish at 4:05 but it's already 4:10. Oh yes sorry etc.

Claude Model

None

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.92

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_