Memory system: stored rules frequently not consulted before acting

Problem

Claude Code's auto-memory system correctly stores user feedback and conventions, but the model frequently fails to consult those memories before taking action. This is especially pronounced with style and formatting conventions that have been corrected multiple times.

Example

A user has a memory entry (feedback_block_comment_spacing.md) that clearly states: trailing blank // lines in multi-line comments should only precede function/class definitions, not const declarations. Despite this rule being stored and indexed in MEMORY.md, the model repeatedly adds the trailing // before const declarations anyway — requiring the user to correct the same mistake across multiple conversations.

Impact

• Users lose trust in the memory system when they have to re-correct behavior that's already been saved
• The value proposition of memory ("so the user doesn't have to offer the same guidance twice") is undermined
• Users end up spending time managing memories that don't reliably influence behavior

Expected behavior

Before emitting code that touches areas covered by stored feedback memories (comment style, naming conventions, test patterns, etc.), the model should actively check relevant memories and apply them. The memories are already loaded into context via MEMORY.md — the issue is that they aren't being used to gate output.

Environment

• Claude Code CLI
• Model: claude-opus-4-6 (1M context)