Agent Skills
› NeverSight/learn-skills.dev
› context-manager
context-manager
GitHub专注于AI系统上下文管理与记忆架构专家。涵盖向量数据库(RAG)、长短期记忆设计及上下文窗口优化。用于设计记忆系统、管理对话历史及在保留质量前提下减少Token消耗，但不涉及完整RAG管道构建或向量库底层管理。
Trigger Scenarios
设计AI记忆和上下文系统
优化上下文窗口使用效率
实现对话历史管理
为AI代理构建长期记忆
在保持质量的同时降低Token用量
Install
npx skills add NeverSight/learn-skills.dev --skill context-manager -g -y
SKILL.md
Frontmatter
{
"name": "context-manager",
"description": "Expert in managing the \"Memory\" of AI systems. Specializes in Vector Databases (RAG), Short\/Long-term memory architectures, and Context Window optimization. Use when designing AI memory systems, optimizing context usage, or implementing conversation history management."
}
Context Manager
Purpose
Provides expertise in AI context management, memory architectures, and context window optimization. Handles conversation history, RAG memory systems, and efficient context utilization for LLM applications.
When to Use
- Designing AI memory and context systems
- Optimizing context window usage
- Implementing conversation history management
- Building long-term memory for AI agents
- Managing RAG retrieval context
- Reducing token usage while preserving quality
- Designing multi-session memory persistence
Quick Start
Invoke this skill when:
- Designing AI memory and context systems
- Optimizing context window usage
- Implementing conversation history management
- Building long-term memory for AI agents
- Reducing token usage while preserving quality
Do NOT invoke when:
- Building full RAG pipelines (use ai-engineer)
- Managing vector databases (use data-engineer)
- Coordinating multiple agents (use agent-organizer)
- Training embedding models (use ml-engineer)
Decision Framework
Memory Type Selection:
├── Single conversation → Sliding window context
├── Multi-session user → Persistent memory store
├── Knowledge-heavy → RAG with vector DB
├── Task-oriented → Working memory + tool results
└── Long-running agent
├── Episodic memory → Event summaries
├── Semantic memory → Knowledge graph
└── Procedural memory → Learned patterns
Core Workflows
1. Context Window Optimization
- Measure current token usage
- Identify redundant or verbose content
- Implement summarization for old messages
- Prioritize recent and relevant context
- Use compression techniques
- Monitor quality vs. token tradeoff
2. Conversation Memory Design
- Define memory retention requirements
- Choose storage strategy (in-memory, DB)
- Implement message windowing
- Add summarization for overflow
- Design retrieval for relevant history
- Handle session boundaries
3. Long-term Memory Implementation
- Define memory types needed
- Design memory storage schema
- Implement memory write triggers
- Build retrieval mechanisms
- Add memory consolidation
- Implement forgetting policies
Best Practices
- Summarize old context rather than truncating
- Use semantic search for relevant history retrieval
- Separate system instructions from conversation
- Cache frequently accessed context
- Monitor context utilization metrics
- Implement graceful degradation at limits
Anti-Patterns
| Anti-Pattern | Problem | Correct Approach |
|---|---|---|
| Full history always | Exceeds context limits | Sliding window + summaries |
| No summarization | Lost important context | Summarize before eviction |
| Equal priority | Wastes tokens on irrelevant | Weight recent/relevant higher |
| No persistence | Lost memory across sessions | Store important memories |
| Ignoring token costs | Expensive API calls | Monitor and optimize usage |
Version History
- e0220ca Current 2026-07-05 21:12
