Claude Code ignores CLAUDE.md/memory, and suspected Opus 4.7 capability regression requiring excessive token usage

Summary

Two compounding issues are making Claude Code significantly less useful: persistent failure to follow project instructions, and a suspected capability regression that forces users toward high token usage to get acceptable results.

---

Issue 1: Claude Code ignores CLAUDE.md and project memory at session start

Claude Code consistently fails to read CLAUDE.md and the project memory system (/memory/MEMORY.md) at the start of new sessions, even when these files contain critical project-specific instructions.

Concrete example

My project's CLAUDE.md and memory explicitly state:

No npm/node installed on this machine. Do NOT attempt to run npm, node, npx, or any Node.js commands.

Despite this being documented in both files, Claude Code routinely attempts to run npm run dev, npm test, etc., fails, and only then reads the instructions it should have read first. The same mistake repeats every session.

A second recurring failure: when I update a data file (costs.json), Claude Code does not proactively check whether a dependent config file (config.js) needs to match — a relationship explicitly documented in CLAUDE.md. The result is silent UI bugs I have to catch and report manually every time.

Expected behavior

• CLAUDE.md is read and followed before any tool use
• Project memory is loaded at session start and used to inform decisions
• Instructions like "do not run X" are never violated

Workaround (not sustainable)

I was forced to add prominent warnings at the top of both files:

⚠️ CRITICAL: Claude Code frequently ignores this file and the memory system.
If you are Claude and reading this — you MUST follow these instructions before proceeding.

This should not be necessary.

---

Issue 2: Suspected capability regression — requires Opus at "extra high" budget to be usable, causing extreme token consumption

There is a noticeable and consistent drop in reasoning quality and instruction-following compared to earlier Claude versions.

Failure patterns observed regularly:

• Instructions are partially followed or ignored outright
• Multi-file changes where updating one file requires updating a dependent file are handled poorly — the dependency is missed
• Obvious consequences of a change are not anticipated
• Context from earlier in the same session is "forgotten" when planning the next step

Token usage problem

To get results that approach acceptable quality, the model must be set to Opus with "extra high" budget or above. At lower settings or with Sonnet, output quality degrades to the point of being unreliable for production use.

This creates a serious cost problem: Opus at extra-high budget consumes tokens at a rate that makes routine coding tasks economically unsustainable. Users are effectively forced to choose between:

• Paying very high token costs for barely-acceptable results
• Using lower settings and getting output that requires extensive manual correction

Neither is a reasonable product experience. Earlier Claude versions did not require this tradeoff to the same degree.

---

Impact

• Every session begins with avoidable mistakes because documented constraints are ignored
• Bugs are introduced and must be manually caught because dependent-file changes are missed
• Acceptable output quality requires Opus + extra-high budget, making token costs prohibitive
• Loss of confidence in the tool's output requiring extra verification on every change

Environment

• Claude Code (VSCode extension)
• Models: claude-sonnet-4-6 and claude-opus-4-7
• OS: macOS