AgentSkillsCN

memory

带有三种 MCP 工具的内存系统：mcp__plugin_kg_kodegen__memory_list_libraries（发现命名空间）、mcp__plugin_kg_kodegen__memory_memorize（存储嵌入）、mcp__plugin_kg_kodegen__memory_recall（语义搜索）。用于按含义存储/检索知识。何时：用户说“记住这个”、“回忆”、“我保存了什么”、“我的笔记”、“查找我的知识”。何时不：文件操作（使用 fs_*）、精确关键词搜索（使用 fs_search）。

SKILL.md
--- frontmatter
name: memory
description: >
  Memory system with three MCP tools: mcp__plugin_kg_kodegen__memory_list_libraries (discover namespaces), mcp__plugin_kg_kodegen__memory_memorize (store with embeddings), mcp__plugin_kg_kodegen__memory_recall (semantic search). Use for storing/retrieving knowledge by meaning.
  WHEN: User says "remember this", "recall", "what did I save", "my notes", "find my knowledge about".
  WHEN NOT: File operations (use fs_*), exact keyword search (use fs_search).

Memory System

Three MCP Tools

1. mcp__plugin_kg_kodegen__memory_list_libraries

Purpose: List all memory library names (namespaces)
Parameters: None - use {}
Returns: Array of library names

2. mcp__plugin_kg_kodegen__memory_memorize

Purpose: Store content with automatic embeddings
Parameters:

  • •library (string, required) - Namespace
  • •content (string, required) - What to store

3. mcp__plugin_kg_kodegen__memory_recall

Purpose: Semantic search by meaning (not keywords)
Parameters:

  • •library (string, required) - Which namespace to search
  • •context (string, required) - Natural language query
  • •limit (number, optional, default: 10) - Max results

When to Use Each

User SaysToolWhy
"remember this", "save this"mcp__plugin_kg_kodegen__memory_memorizeStore new knowledge
"what did I save about X"mcp__plugin_kg_kodegen__memory_recallFind by semantic meaning
"what libraries exist"mcp__plugin_kg_kodegen__memory_list_librariesDiscover namespaces

Key Concepts

Semantic Search: Finds by meaning, not exact words. "authentication flow" matches "login process", "OAuth", "user sign-in" because semantically related.

Libraries as Namespaces: Each library is isolated. Use patterns like:

  • •project-{name} for project knowledge
  • •snippets-{lang} for code snippets
  • •notes-{context} for contextual notes

Content Best Practices:

  • •Include context, not just raw data
  • •One concept per memory
  • •Self-contained (understandable alone)
  • •Natural language with searchable terms

Common Workflows

Store:

json
{"library": "project-notes", "content": "Authentication uses OAuth2 with Google. Tokens in HTTP-only cookies, 24h timeout."}

Discover then Search:

json
// 1. List libraries
{}

// 2. Search specific library
{"library": "project-notes", "context": "authentication decisions", "limit": 5}

Cross-library Search: Use mcp__plugin_kg_kodegen__memory_list_libraries, then mcp__plugin_kg_kodegen__memory_recall from each relevant library.

Query Tips

✅ DO: "async HTTP requests with retry and timeout"
❌ DON'T: "http" or "retry"

Use natural language, describe what you're looking for. More context = better results.

Remember

  • •Libraries created on first use (no pre-creation needed)
  • •Memories are additive (storing doesn't replace)
  • •Embeddings are automatic
  • •Results ranked by semantic similarity