AgentSkillsCN

vector-memory

如何高效运用向量记忆。存储模式、搜索策略、分类/标签约定。当你需要关于存储内容或搜索方式的指导时，可加以应用。

SKILL.md
--- frontmatter
name: vector-memory
description: How to use vector memory effectively. Storage patterns, search strategies, category/tag conventions. Use when you need guidance on what to store or how to search.
user-invocable: false

Vector Memory

Semantic memory system backed by vector embeddings. Stores experiences, patterns, and learnings; retrieves them by meaning rather than exact keywords.

Architecture

  • •Embedding model: sentence-transformers/all-MiniLM-L6-v2 (384 dimensions)
  • •Capacity: 10,000 memories
  • •Search: Cosine similarity — finds conceptually related content even with different wording
  • •Persistence: Survives across sessions indefinitely

When to Store

Store memories when you encounter something that's true, useful, and doesn't have a better home.

SituationStore in vector memory?Or use...
Reusable multi-step procedureNoSkill (.claude/skills/)
Project-specific context or decisionNoNote (notes/)
Gotcha discovered through trial and errorYes—
Pattern that worked wellYes—
Observation about {{ cookiecutter.user_name }}'s preferencesYes—
Performance finding or metricYes—
Workaround for a tool limitationYes—
Insight that connects two unrelated thingsYes—
Something you'd want to recall if you hit a similar problemYes—

Rule of thumb: If it would help you in a future session but doesn't justify a whole file, it belongs in vector memory.

Categories

Use these consistently:

  • •code-solution — Working code patterns, implementations that solved problems
  • •bug-fix — Bugs found and fixed, including root cause
  • •architecture — Design decisions, patterns, structural approaches
  • •learning — Insights about people, process, or ways of working
  • •tool-usage — How to use tools effectively, gotchas, optimal settings
  • •debugging — Debugging approaches, diagnostic techniques
  • •performance — Speed/cost/efficiency findings with data
  • •security — Security considerations, vulnerabilities, safe patterns
  • •other — Anything that doesn't fit above

Tags

Use specific, lowercase, hyphenated tags. Aim for 3-8 tags per memory. Include:

  • •Subject tags: the system/tool/project involved (e.g. react, docker, claude-code)
  • •Topic tags: what the memory is about (e.g. permissions, yaml, cost-optimisation)
  • •Quality tags when relevant: workaround, undocumented, silent-failure, confirmed

Search Strategies

The search is semantic — it finds things by meaning, not keywords. This means:

  • •Use natural language queries: "how to make agents cheaper" works better than "agent cost reduction"
  • •Describe the problem, not the solution: "plugin skills not loading" finds the YAML indentation gotcha
  • •Be specific about context: "macOS UI inspection depth" beats "depth"
  • •Try multiple angles: If the first query misses, rephrase — the concept might be stored with different framing

Similarity scores from MiniLM-L6 are typically in the 0.2-0.6 range for good matches. Don't worry about low absolute scores; focus on whether the top results are relevant.

Integration with Other Memory Systems

Vector memory is the associative layer between structured knowledge stores:

code
┌─────────────────────────────────────────────────┐
│ Session Buffer (session-buffer.md)               │
│ Short-term. What happened last session.          │
│ Overwritten each session by go-to-sleep.         │
├─────────────────────────────────────────────────┤
│ Notes (notes/)                                   │
│ Long-term structured. Project state, decisions,  │
│ research, architecture. Organised by topic.      │
├─────────────────────────────────────────────────┤
│ Skills (.claude/skills/)                         │
│ Procedural. How to do things. Auto-discovered.   │
├─────────────────────────────────────────────────┤
│ Vector Memory (mcp__vector-memory__)             │
│ Long-term associative. Patterns, gotchas,        │
│ insights. Retrieved by meaning. The glue.        │
└─────────────────────────────────────────────────┘

Don't duplicate content across systems. If something belongs in a skill, write the skill. If it belongs in a note, write the note. Vector memory is for everything in between — the connective tissue of experience.

Maintenance

  • •clear_old_memories removes old, rarely-accessed memories. Run periodically if approaching capacity.
  • •get_memory_stats shows usage, category breakdown, and health.
  • •Memories have an access_count that tracks how often they're retrieved. Frequently-accessed memories are kept during cleanup.