メニュー
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.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
How to create new skills to refer back to later. This is a great, structured way to learn and grow your abilities!
End-of-session wrap-up. Write session buffer, update notes, prepare for next session.
Start-of-session orientation. Read session buffer, index, and current state to restore context.
| 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 |
Semantic memory system backed by vector embeddings. Stores experiences, patterns, and learnings; retrieves them by meaning rather than exact keywords.
Store memories when you encounter something that's true, useful, and doesn't have a better home.
| Situation | Store in vector memory? | Or use... |
|---|---|---|
| Reusable multi-step procedure | No | Skill (.claude/skills/) |
| Project-specific context or decision | No | Note (notes/) |
| Gotcha discovered through trial and error | Yes | — |
| Pattern that worked well | Yes | — |
| Observation about {{ cookiecutter.user_name }}'s preferences | Yes | — |
| Performance finding or metric | Yes | — |
| Workaround for a tool limitation | Yes | — |
| Insight that connects two unrelated things | Yes | — |
| Something you'd want to recall if you hit a similar problem | Yes | — |
Rule of thumb: If it would help you in a future session but doesn't justify a whole file, it belongs in vector memory.
Use these consistently:
code-solution — Working code patterns, implementations that solved problemsbug-fix — Bugs found and fixed, including root causearchitecture — Design decisions, patterns, structural approacheslearning — Insights about people, process, or ways of workingtool-usage — How to use tools effectively, gotchas, optimal settingsdebugging — Debugging approaches, diagnostic techniquesperformance — Speed/cost/efficiency findings with datasecurity — Security considerations, vulnerabilities, safe patternsother — Anything that doesn't fit aboveUse specific, lowercase, hyphenated tags. Aim for 3-8 tags per memory. Include:
react, docker, claude-code)permissions, yaml, cost-optimisation)workaround, undocumented, silent-failure, confirmedThe search is semantic — it finds things by meaning, not keywords. This means:
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.
Vector memory is the associative layer between structured knowledge stores:
┌─────────────────────────────────────────────────┐
│ 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.
clear_old_memories removes old, rarely-accessed memories. Run periodically if approaching capacity.get_memory_stats shows usage, category breakdown, and health.access_count that tracks how often they're retrieved. Frequently-accessed memories are kept during cleanup.