AgentSkillsCN

memory

使用 minimem 搜索并管理记忆。当用户希望回溯过往的上下文、决策、笔记或已存储的信息时，可随时调用此功能。此外，当需要存储新信息以便日后检索时，亦可选用 minimem。

SKILL.md
--- frontmatter
name: memory
description: Search and manage memories using minimem. Use when the user wants to recall previous context, decisions, notes, or stored information. Also use when storing new information for later retrieval.

Memory Skill

Use the minimem memory system to store and retrieve persistent context.

When to Use This Skill

  • •User asks to "remember" or "store" something
  • •User asks to "recall", "find", or "search for" previous context
  • •User references past decisions, notes, or information
  • •You need to persist information across sessions
  • •User mentions "memory" or "memories"

Available MCP Tools

memory_search (two-phase search)

code
memory_search(query, maxResults?, minScore?, directories?, detail?, type?)
  • •query: Natural language search query
  • •maxResults: Maximum results (default: 10)
  • •minScore: Minimum relevance 0-1 (default: 0.3)
  • •directories: Filter to specific memory directories
  • •detail: "compact" (default) for lightweight index, "full" for complete snippets
  • •type: Filter by observation type (decision, bugfix, feature, discovery, context, note)

memory_get_details

code
memory_get_details(results, directories?)
  • •results: Array of { path, startLine, endLine } from compact search results
  • •directories: Filter to specific memory directories

Searching Memories (Two-Phase Workflow)

When the user wants to recall information:

  1. •Use memory_search with a descriptive query (defaults to compact results)
  2. •Review the compact index — each result shows path, score, and short preview
  3. •Use memory_get_details to fetch full text for the most relevant results
  4. •Synthesize and present the information, citing source files

This two-phase approach saves tokens by only fetching details for results that matter.

For quick lookups where you need all details immediately, use detail: "full".

Filtering by Type

Use the type parameter to narrow results:

  • •memory_search(query: "API design", type: "decision") — only decisions
  • •memory_search(query: "login bug", type: "bugfix") — only bug fixes

Storing Memories

To store new information, use filesystem tools to write to memory files:

  1. •Quick notes → Append to memory/YYYY-MM-DD.md (today's date)
  2. •Decisions → Add to MEMORY.md under appropriate heading
  3. •Topic-specific → Create/update memory/<topic>.md

Storage Format

markdown
### YYYY-MM-DD HH:MM
<!-- type: decision -->
<content to remember>

Observation Types

Add a type comment to categorize entries for filtered search:

  • •<!-- type: decision --> — architectural or design decisions
  • •<!-- type: bugfix --> — bug fixes and root causes
  • •<!-- type: feature --> — new feature implementations
  • •<!-- type: discovery --> — learned facts, insights
  • •<!-- type: context --> — project context, setup notes
  • •<!-- type: note --> — general notes (default)

Privacy

Wrap sensitive content in <private> tags to exclude it from search indexing:

markdown
API endpoint: api.example.com/v2

<private>
API_KEY=sk-abc123
DB_PASSWORD=hunter2
</private>

Best Practices

  1. •Use specific, searchable terms when storing
  2. •Include context: dates, people, project names
  3. •Add type comments for important entries (decisions, bugfixes)
  4. •Use <private> tags for credentials and secrets
  5. •Start with compact search, then fetch details as needed
  6. •When searching, try multiple query variations if needed