スキル agent-memory-systems
📦

agent-memory-systems

安全

Design Agent Memory Systems

Build intelligent memory architectures for AI agents that enable long-term learning and context-aware responses. This skill provides patterns for implementing short-term, long-term, and episodic memory that improves agent performance over time.

対応: Claude Codex Code(CC)
🥉 74 ブロンズ
1

スキルZIPをダウンロード

2

Claudeでアップロード

設定 → 機能 → スキル → スキルをアップロードへ移動

3

オンにして利用開始

テストする

「agent-memory-systems」を使用しています。 Design memory for a code review agent that learns from past reviews

期待される結果:

A code review agent should use episodic memory to store individual review sessions, semantic memory for learned coding patterns, and procedural memory for review workflows. Implement retrieval with temporal weighting to prioritize recent relevant reviews while avoiding stale context.

「agent-memory-systems」を使用しています。 What memory type should store user preferences?

期待される結果:

User preferences should be stored in semantic memory since they represent persistent learned knowledge about the user. Use episodic memory to track how preferences evolved over time. Apply metadata filtering to retrieve user-specific memories efficiently.

セキュリティ監査

安全
v1 • 2/24/2026

This skill is a pure documentation resource about agent memory architecture. Static analysis flagged apparent security patterns (backtick syntax, cryptographic keywords), but manual review confirms these are false positives. The backticks are markdown code formatting in a skill list, not shell execution. No actual code, scripts, network calls, or cryptographic operations exist. This is safe for publication.

1
スキャンされたファイル
72
解析された行数
2
検出結果
1
総監査数
中リスクの問題 (1)
False Positive: External Commands Flag
Static scanner flagged 'Ruby/shell backtick execution' at SKILL.md:68. This is markdown code formatting for a list of related skills: `autonomous-agents`, `multi-agent-orchestration`, etc. No shell execution occurs.
低リスクの問題 (1)
False Positive: Weak Cryptographic Algorithm
Static scanner flagged 'weak cryptographic algorithm' at SKILL.md:3 and SKILL.md:71. These are YAML frontmatter keywords and regular text about 'execution' - no cryptographic code exists.
監査者: claude

品質スコア

38
アーキテクチャ
100
保守性
87
コンテンツ
50
コミュニティ
98
セキュリティ
91
仕様準拠

作れるもの

AI Application Developer Building Conversational Agents

Design memory systems that allow chatbots to remember user preferences and conversation history across sessions, enabling more personalized and context-aware interactions.

DevOps Engineer Building Autonomous Agents

Implement long-term memory for agents that automate infrastructure tasks, allowing them to learn from previous deployments and avoid repeating mistakes.

Research Engineer Prototyping Agent Architectures

Explore different memory system designs and retrieval strategies to find optimal configurations for specific agent use cases.

これらのプロンプトを試す

Design Memory Architecture
Design a memory system for an AI agent that handles customer support conversations. The agent needs to remember user preferences, past issues, and resolution history. Specify the memory types, storage approach, and retrieval strategy.
Evaluate Chunking Strategy
Review this document chunking strategy for a knowledge base: [describe strategy]. What are the strengths and weaknesses? How would you improve retrieval accuracy?
Fix Memory Anti-Pattern
Our agent memory system is storing every conversation indefinitely. This is causing slow retrieval and high storage costs. What memory management approach should we use instead?
Select Vector Store
Compare vector database options for a production agent system handling 10 million embeddings with sub-100ms retrieval requirements. Consider Pinecone, Weaviate, Chroma, and pgvector.

ベストプラクティス

  • Always filter by metadata (time, user, topic) before semantic retrieval to reduce noise and improve relevance
  • Implement memory decay or importance scoring to prevent old or irrelevant memories from polluting context
  • Budget tokens strategically across memory types - prioritize recent, high-relevance memories in limited context windows

回避

  • Storing everything forever without eviction policies - leads to retrieval pollution and increased costs
  • Using a single memory type for all data types - different information needs different storage approaches
  • Chunking documents without testing retrieval quality - chunks that look good semantically may not retrieve well

よくある質問

What is the difference between short-term and long-term memory?
Short-term memory (working memory) lives in the context window and is available during a single conversation. Long-term memory persists across sessions and requires a vector store or database for retrieval.
When should I use episodic vs semantic memory?
Use episodic memory for specific events or interactions (a conversation, a task completion). Use semantic memory for learned facts and patterns (user preferences, general knowledge).
How do I prevent memory retrieval from returning irrelevant results?
Apply metadata filtering before semantic search (filter by time, user, topic). Use contextual chunking that preserves surrounding context. Implement importance scoring to boost recent or critical memories.
What chunk size should I use for document storage?
There is no universal answer. Test different sizes with your actual retrieval queries. Smaller chunks (256-512 tokens) work well for precise matching, larger chunks (1024+) preserve more context but may introduce noise.
Can Claude Code execute memory system code?
Claude Code can help design and explain memory systems but cannot execute them directly. This skill provides architectural guidance that developers implement in their own code.
How do I handle conflicting memories?
Detect conflicts on storage by comparing new memories against existing ones. Store conflict markers or use versioning. At retrieval time, surface the most recent or highest-confidence memory and flag conflicts to the user.

開発者の詳細

ファイル構成

📄 SKILL.md