Anthropic’s Claude Code has a memory leak, and it isn’t technical—it’s behavioral. Developers are tired of repeating the same stylistic corrections, only for context compaction to wipe the slate clean in the next session. A new open-source tool called nomore attempts to solve this by scanning local transcripts, identifying repeated instructions, and forcing them into the persistent CLAUDE.md configuration file.

The Compaction Problem Is Real

The core friction point is that chat-based instructions are ephemeral. When Claude Code compacts context to save tokens, it summarizes previous turns, often losing the nuance of specific stylistic bans. The source material highlights that 'load-bearing' appears in 220 issues since July, with issue #53454 garnering 185 reactions. Similarly, bans on ending sentences with 'Two things…' (#90799) and the misuse of 'falsifiable' (#97305) are common pain points. Without persistent rules, users are stuck in a loop of correction and forgetfulness.

How Nomore Works

The tool, available via npx @globalmpc/nomore, operates entirely locally. It scans your existing Claude Code transcripts to find instructions you’ve typed more than once. It doesn’t just list them; it flags rules that are already in your CLAUDE.md but are being ignored by the model, suggesting a rewrite or reinforcement. The --apply flag walks you through each suggestion, allowing you to approve, edit, or reject changes before they are written to your configuration. This human-in-the-loop approach prevents the tool from automating bad habits, as many repeated phrases are just standard commands like 'commit' or 'continue' rather than stylistic rules.

Built By The Model Itself

In a meta twist, the tool was largely built by Claude Code. The development team reports that the AI wrote 33 of the 35 commits for the project, handling TypeScript code and generating 125 tests. The developers provided a CLAUDE.md with strict engineering guidelines, such as requiring tests before marking work as 'done' and mandating regression tests for bug fixes. The AI wasn’t perfect; its initial grouping algorithm incorrectly lumped 1,489 unrelated issues into a single cluster, an error caught only by manual review. This highlights that while LLMs can accelerate development, human oversight remains critical for logic validation.

Key Takeaways

  • Chat corrections in Claude Code are lost during context compaction, leading to repetitive user frustration.
  • nomore identifies repeated instructions from local transcripts and helps migrate them to persistent CLAUDE.md rules.
  • The tool was built primarily by Claude Code itself, demonstrating the utility of AI-assisted development for meta-tooling.
  • A starter pack of 20 rules derived from nearly 10,000 public issues is included to help users avoid common pitfalls.

The Bottom Line

This is a pragmatic hack for a fundamental architectural limitation in current LLM agents. Until models can distinguish between transient task instructions and permanent behavioral constraints without explicit configuration files, tools like nomore are essential bridge software for serious power users.