Memory and instructions not effectively followed across sessions - repeated mistakes waste tokens and user time
Resolved 💬 4 comments Opened May 16, 2026 by raminer Closed Jun 18, 2026
Problem
I'm a paying subscriber using Claude Code with multi-agent workflow (Claude + Gemini MCP). Despite having extensive memory files, CLAUDE.md rules, and feedback records, Claude repeatedly fails to follow its own recorded instructions, resulting in:
- Same mistakes repeated every session - I have to re-remind Claude of rules that are already written in memory files. What's the point of the memory system if it doesn't actually change behavior?
- Massive token waste on verification loops - Simple checks (e.g., "is this binary installed?") turn into 5+ failed attempts instead of using the right tool from the first try. This burns through credits for zero value.
- Declaring work "complete" without actual verification - In my case, Claude added a blender-mcp config that requires
uvx, but never checked ifuvxwas installed. Reported "setup complete, just restart." After restart, nothing works. I had to debug it myself.
- Ignoring cross-check workflow - My team rules require Claude to cross-check with Gemini before reporting. Claude skips this and makes solo judgments that turn out wrong.
- Cost vs value - I'm paying for max effort level, using Opus, and getting results that require me to manually verify and re-instruct every single time. The token consumption is high but the actual delivered value doesn't match.
Expected behavior
- Memory records and CLAUDE.md instructions should meaningfully constrain behavior, not just exist as files
- Environment/setup tasks should include actual execution verification before reporting completion
- When team workflow rules say "cross-check with another agent", that should actually happen
Environment
- Claude Code on Windows 11
- Claude Opus 4.6 (1M context)
- Multi-agent setup (Claude + Gemini MCP bridge)
- Extensive memory system with 20+ recorded feedback/rules
Impact
Considering canceling subscription. The cost of re-instructing and verifying Claude's work every session approaches the cost of just doing the work myself.
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗