Claude ignores memorized instructions and repeats corrected mistakes

Bug Description

Claude Code's persistent memory system stores user corrections and explicit instructions for future behavior. However, Claude repeatedly ignores these memorized instructions and makes the exact same mistakes that were already corrected and saved to memory.

The pattern:

1. User corrects Claude's behavior
2. Correction is saved to memory with clear rules, examples of correct/incorrect usage, and reasoning
3. In a later (or even the same) conversation, Claude ignores the memory and repeats the original mistake
4. When the first attempt fails, Claude tries alternative approaches that are also explicitly forbidden in the same memory entry
5. User has to correct Claude again — the exact same correction that's already in memory

Example

A memory entry explicitly states:

• Correct syntax: command "arg"
• Wrong syntax: command -flag "arg" (duplicates a flag already in the alias)
• Also wrong: bypassing the alias entirely with the raw command

Claude proceeds to use the wrong syntax, fails, then falls back to the raw command — both explicitly documented as incorrect in loaded memory.

Expected Behavior

When memory entries contain explicit do/don't instructions with examples, Claude should follow them on the first attempt. The entire point of persistent memory is to avoid repeating the same corrections across conversations.

Actual Behavior

Memory entries are loaded into context but functionally ignored. Claude falls back to its default behavior patterns, treating memorized user corrections as suggestions rather than constraints.

Impact

This completely undermines trust in the memory system. Users invest effort in correcting behavior and verifying it's saved, only to discover that Claude will repeat the same mistakes anyway. The memory system becomes write-only — data goes in but doesn't reliably influence behavior.