Auto memory (MEMORY.md) is not useful and clutters context

Summary

The auto memory feature (MEMORY.md in .claude/projects/) is not useful in practice and actively gets in the way.

Problems

1. No practical value: The information stored in MEMORY.md has never been useful for actual task execution. Project-specific rules and patterns are better managed through dedicated instruction systems (e.g., MCP-based instruction tools, CLAUDE.md).

1. Clutters context: MEMORY.md content is always loaded into conversation context, consuming tokens that could be used for actual work. This is especially problematic in long sessions working on complex tasks.

1. Duplicates other systems: Projects that use structured instruction management (e.g., interactive-instruction MCP tools) already have a proper place for rules and knowledge. MEMORY.md becomes a redundant, less authoritative copy that can conflict with the canonical source.

1. AI writes to it reflexively: The agent tends to write findings to MEMORY.md as a default behavior, even when the information belongs in a project's proper documentation system. This creates maintenance burden and confusion about where the source of truth lives.

Suggestion

Consider making auto memory opt-in rather than opt-out, or provide a clear way to disable it per-project.

Environment

• Claude Code CLI
• macOS