---
id: 8e2d1904-24e0-41df-95f2-2edc4407b9f3
type: solution
title: "Converted MemoryGraph MCP to Claude Code Skill"
tags: [pai, claude-code, memorygraph, skills, architecture, migration]
importance: 0.9
confidence: 0.8
created: "2025-12-07T06:50:12.011310+00:00"
updated: "2026-03-01T06:21:14.622001+00:00"
relations:
  - target: ebc3e5af-b3d1-486b-9751-8ece841aa93f
    type: RELATED_TO
    direction: outgoing
    strength: 0.5
    context: "Both part of PAI skills infrastructure improvements in same session"
  - target: 089f4cd8-4b2b-49a5-aac1-af2811b8ea7d
    type: RELATED_TO
    direction: incoming
    strength: 0.8
    edge_id: 34b6ee00-1113-479b-8e28-e39f347c6bc0
  - target: 8023fae6-eec1-4c93-bac8-1bd85dd61bae
    type: BUILDS_ON
    direction: incoming
    strength: 0.74
    edge_id: b4b234af-a7be-4144-b965-75642070199f
  - target: dfe48a70-b8ab-4bb5-a3d8-6116fa498168
    type: RELATED_TO
    direction: incoming
    strength: 0.76
    edge_id: 51e04dfd-be31-4d5e-ab33-04ce2858a18f
---

Converted MemoryGraph from MCP server to Claude Code skill to reduce token cost. MCP tools were always in context (~45 tools eating tokens every turn). Skill loads on-demand only when triggered.

Implementation:
- Created ~/.claude/skills/memorygraph/ with client.py (620 lines), SKILL.md, SCHEMA.md, feature.json
- Direct SQLite access to ~/.memorygraph/memory.db bypassing MCP protocol
- JSON-based search instead of FTS (existing FTS index wasn't populated)
- CLI interface: store, recall, get, relate, search, update, delete, related, stats, recent

All 11 core operations implemented:
- store_memory, recall_memories, get_memory, create_relationship (high priority)
- search_memories, update_memory, delete_memory, get_related_memories (medium)
- get_memory_statistics, get_recent_activity (low)

Updated ~/.claude/CLAUDE.md Memory Protocol to use CLI commands instead of MCP tools.
Documentation saved to NoteDiscovery at reference/skills/memorygraph.