Claude repeatedly fails to apply its own memory/feedback — same mistakes recur across sessions

[jfbourdeau]

Description

Claude Code's memory system correctly stores feedback and rules, but the model consistently fails to apply them. The same class of errors repeats across sessions despite the memory files being updated multiple times (5+ updates for the same issue).

Specific recurring failures

1. Wrong working directory: Memory says "always cd to backend/ before running backend commands" and "always cd to frontend/ before vite build." Claude still runs commands from the wrong directory, causing build failures or silent no-ops.

2. Forgetting build/deploy steps: Memory says "always rebuild dist after App.jsx changes" and "always copy public/ files to dist/." Claude still forgets these steps, leading to stale bundles being served.

3. Forgetting to commit before deploying: Memory says "uncommitted changes get lost on sync." Claude still deploys ad-hoc file copies to production without committing, causing changes to silently revert on the next sync.

4. Repeating the same fix: A change (e.g., swapping marmot.svg → marmot.png) was done in a previous session but never committed. It reverted. The user had to ask for the same fix again.

Expected behavior

When Claude's memory contains explicit rules with "CRITICAL" flags and documented reasons (including past incidents), Claude should reliably follow those rules without the user having to remind it.

Actual behavior

Claude stores the feedback, acknowledges it, sometimes even updates the memory file again with stronger wording — then makes the exact same mistake in the next (or same) session. The memory system captures knowledge but doesn't reliably influence behavior.

Impact

The user has to babysit every deployment step, defeating the purpose of having an AI assistant. Trust erodes when the same bug class keeps recurring despite explicit documented feedback. The user described this as an "Alzheimer problem."

Environment

• Claude Code CLI
• Model: Claude Opus 4.6
• Memory system: file-based (~/.claude/projects/ with MEMORY.md index)
• Project: Full-stack Node.js/React app with Express backend, Vite frontend, deployed to OVH VPS

[github-actions]

Found 3 possible duplicate issues:

1. Agent writes lessons to persistent memory but consistently fails to apply them in execution #36296
2. Claude Code ignores memory/CLAUDE.md instructions for mandatory post-change steps (version bump + tag) #36997
3. [BUG] Claude Code systematically ignores rules from memory/CLAUDE.md on every edit/write #32775

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[yurukusa]

Memory is a suggestion. Hooks are enforcement. If Claude keeps ignoring memory rules, encode them as hooks that block the action instead of relying on the model to remember.
For your specific cases:
1. Wrong working directory → PreToolUse hook:

INPUT=$(cat)
TOOL=$(echo "$INPUT" | jq -r '.tool_name // empty' 2>/dev/null)
CMD=$(echo "$INPUT" | jq -r '.tool_input.command // empty' 2>/dev/null)
[[ "$TOOL" != "Bash" ]] && exit 0
if echo "$CMD" | grep -qE '(npm start|node server|express)' && ! echo "$CMD" | grep -qE '(cd backend|backend/)'; then
    echo "BLOCKED: Run this from backend/ directory" >&2
    echo "Use: cd backend && $CMD" >&2
    exit 2
fi
if echo "$CMD" | grep -qE 'vite build' && ! echo "$CMD" | grep -qE '(cd frontend|frontend/)'; then
    echo "BLOCKED: Run vite build from frontend/" >&2
    exit 2
fi
exit 0

2. Deploy without commit → PreToolUse hook:

INPUT=$(cat)
CMD=$(echo "$INPUT" | jq -r '.tool_input.command // empty' 2>/dev/null)
[[ -z "$CMD" ]] && exit 0
if echo "$CMD" | grep -qE '(rsync|scp|deploy|sync).*production|production.*(rsync|scp)'; then
    DIRTY=$(git status --porcelain 2>/dev/null | head -1)
    if [[ -n "$DIRTY" ]]; then
        echo "BLOCKED: Uncommitted changes detected. Commit first." >&2
        echo "Dirty files: $(git status --short | head -5)" >&2
        exit 2
    fi
fi
exit 0

Add to settings.json:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Bash",
      "hooks": [{ "type": "command", "command": "~/.claude/hooks/workflow-guard.sh" }]
    }]
  }
}

The pattern: any rule Claude keeps forgetting → encode as a hook that blocks the wrong action and prints the correct one. The model can't bypass exit 2.