Memory Quality MCP

A Claude Code MCP plugin that audits and cleans up your AI memory.

中文版 · Report a Bug · Request a Feature

---

---

Claude Code v2.1.59+ automatically saves memories from your conversations. Over time, your memory store accumulates:

• Stale memories — "working on Project X this week" from months ago
• Junk memories — offhand remarks treated as permanent facts
• Conflicting memories — "prefers detailed comments" and "keep code clean and minimal" coexisting
• Misrecorded memories — AI over-interpreted a one-time comment as a fixed habit

This plugin audits your memory store with a 4-dimension quality score (Importance / Recency / Credibility / Accuracy) and gives you actionable cleanup recommendations — with a visual dashboard.

---

Requirements

• Claude Code v2.1.59+ (run claude --version to check)
• Python 3.10+
• LLM API Key — OpenAI, Kimi, MiniMax, or Anthropic (any one will do)

---

Installation

Option 1: uvx (recommended — no manual install)

Add to your Claude Code MCP config (~/.claude/settings.json or your project's .claude/settings.json):

{
  "mcpServers": {
    "memory-quality": {
      "command": "uvx",
      "args": ["memory-quality-mcp"],
      "env": {
        "OPENAI_API_KEY": "your-key-here"
      }
    }
  }
}

Option 2: pip

pip install memory-quality-mcp

Then in your MCP config:

{
  "mcpServers": {
    "memory-quality": {
      "command": "memory-quality-mcp",
      "env": {
        "MINIMAX_API_KEY": "your-key-here"
      }
    }
  }
}

---

Configure your LLM

Set one of these environment variables in the env field above:

Provider	Env Variable	Default Model
OpenAI	OPENAI_API_KEY	gpt-4o-mini
Kimi	KIMI_API_KEY	moonshot-v1-8k
MiniMax	MINIMAX_API_KEY	MiniMax-M2.5
Anthropic	ANTHROPIC_API_KEY	claude-haiku-4-5

The plugin auto-detects which provider to use based on whichever key is set.

For advanced config (custom model, thresholds), edit ~/.memory-quality-mcp/config.yaml — generated automatically on first run.

---

Usage

After setup, restart Claude Code and talk to it naturally:

1. Try the demo first (no memories needed)

Open the memory dashboard in demo mode

Claude calls memory_dashboard(demo=True) → opens a browser page with example data so you can see what the tool does before running it on your real memories.

2. Quick health check (no LLM cost)

Check my memory store health

Returns: total memories across all projects, stale count, index usage, estimated LLM calls for a full report.

3. Full quality analysis

Run a detailed memory quality analysis

Scores every memory on 4 dimensions, detects conflicts, flags rule violations. Results are cached — cleanup and dashboard reuse them without calling the LLM again.

Cost estimate: 8–9 LLM calls for 50 memories ($0.01 with gpt-4o-mini)

4. Open visual dashboard

Open the memory health dashboard

Opens a local HTML page in your browser:

• Health score ring (0–100)
• Summary stats (kept / review / delete)
• Collapsible memory list — click any item to expand score details

5. Clean up

Clean up the memories marked for deletion

Claude previews with memory_cleanup(dry_run=True) first, then executes on confirmation.

Safety guarantees:

• Preview before every deletion
• Auto-backup to .trash/<timestamp>/ before deleting
• Never silent-deletes anything

6. Score a single memory (debug)

Score this memory: "User always codes late at night"

7. Analyze a specific project

Analyze memories for ~/my-project only

---

How scoring works

Dimension	Weight	What it measures
Importance	40%	How useful is this for future conversations?
Recency	25%	Is this information still accurate?
Credibility	15%	Is there a clear source (user stated it explicitly)?
Accuracy	20%	Did the AI record it faithfully, without over-interpreting?

Score > 3.5 → Keep ｜ 2.5–3.5 → Review ｜ < 2.5 → Delete

---

Typical workflow

You: Check my memory health

Claude: [memory_audit()]
        47 memories across 3 projects
        - Possibly stale: 8
        - Project memories past threshold: 3
        - MEMORY.md usage: 23% (46/200 lines)
        Estimated ~9 LLM calls for a full report.

You: Run the full analysis

Claude: [memory_report()]
        47 memories | 🗑 delete 8 | 🔄 review 5 | ✅ keep 34

        ⚡ 1 conflict found
        - 🔴 feedback_comments_a.md × feedback_comments_b.md
          One says "detailed comments", the other says "minimal comments"

        🗑 Suggested deletions (8)
        - project_q1_plan.md [project] · 120 days ago
          Score 1.5 · project memory past 90-day threshold

        Report cached — cleanup won't need to re-analyze

You: Open the dashboard

Claude: [memory_dashboard()]
        ✅ Dashboard opened in browser

You: Clean those up

Claude: [memory_cleanup(dry_run=True)]
        🔍 Preview (nothing deleted yet)
        8 memories will be removed: project_q1_plan.md ...

You: Confirm

Claude: [memory_cleanup(dry_run=False)]
        ✅ Cleaned up 8 memories
        Backup: ~/.claude/.../memory/.trash/20260405_143022/
        MEMORY.md index updated

---

FAQ

Q: "No memory files found"

Requires Claude Code v2.1.59+. Check:

claude --version              # must be >= v2.1.59
ls ~/.claude/projects/        # should list your projects

If the version is new enough but no files exist yet, Claude hasn't decided anything is worth remembering yet. Keep using Claude Code normally — files will appear within a few sessions.

Want to see the tool in action first? Run memory_dashboard(demo=True).

Q: Are the scores accurate?

Scores are suggestions, not commands. You make the final call. Every deletion requires confirmation and is backed up. If you find consistent scoring errors, please open a Wrong Score issue — each report directly improves the model.

Q: Which LLMs are supported?

Any OpenAI-compatible API. Built-in presets: OpenAI, Kimi, MiniMax, Anthropic. Custom providers via MEMORY_QUALITY_BASE_URL environment variable.

Q: Will it delete something important?

No. Every operation: ① shows a preview ② requires explicit confirmation ③ auto-backs up to .trash/ before deleting.

Q: Where are my memory files?

ls ~/.claude/projects/*/memory/

Or type /memory inside Claude Code to browse and edit them directly.

---

Contributing

Found a scoring error? Tell us — your feedback directly calibrates the model.

Bug or feature idea? Open an issue.

---

License

MIT