Description

管理多个 AI Agent 的长期记忆与文件，通过向量和时间双重过滤，实现高效语义检索和减少 LLM 调用。

README (SKILL.md)

Context Manager Skill

Name: Context Manager
Author: janussilence

长上下文管理器技能。为多个 AI Agent 提供智能的长期记忆管理，通过向量索引和时间轴双重命中判断，大幅减少 LLM 调用次数。

🎯 适用场景

当用户需要：

管理多个 Agent 的文件和上下文
在大型项目中快速检索相关信息
减少 LLM 调用以降低成本
建立长期记忆系统
在多个 Agent 之间共享上下文
基于语义相似度搜索文件
按时间维度过滤和检索内容

📋 功能清单

Agent 管理

创建新的 Agent
查看 Agent 列表
获取 Agent 详情
删除 Agent

文件管理

为 Agent 创建文件
获取文件内容
列出 Agent 的所有文件
删除文件

智能检索

语义向量搜索
时间轴过滤
双重命中判断（向量+时间）
智能调取详情
LLM 集成（可选）

🚀 使用方法

基本用法

创建 Agent

为项目 "文档编写" 创建一个 Agent

创建文件

将以下内容保存到 Agent "agent001" 的文件 "设计文档.md" 中：
[文件内容]

查询

在 Agent "agent001" 中搜索 "深度学习算法"

查询（带详情）

查找 "微服务架构设计" 相关内容，并获取详情

列出文件

显示 Agent "agent001" 的所有文件

高级用法

带标签的文件

创建一个关于 "机器学习" 的文件，标签为 "research, ai, ml"

跨 Agent 搜索

在所有 Agent 中搜索 "Python 编程"

强制 LLM 调用

用 LLM 分析 "如何优化数据库查询" 相关内容

🔧 配置

环境变量（可选）

# OpenAI API Key（用于 LLM 功能）
export OPENAI_API_KEY="your-api-key"

# Redis 配置（用于缓存）
export CONTEXT_MANAGER_REDIS_HOST="localhost"
export CONTEXT_MANAGER_REDIS_PORT=6379

配置参数（可选）

在 config.py 中可调整：

vector_threshold: 向量相似度阈值（默认 0.7）
time_threshold: 时间相关性阈值（默认 0.5）
combined_threshold: 综合评分阈值（默认 0.6）
cache_ttl: 缓存过期时间（默认 3600 秒）

📊 工作原理

文件索引流程

接收文件内容
生成向量嵌入
存储到向量数据库
提取时间轴信息
更新索引

智能检索流程

向量搜索找到相关文件
时间轴过滤
双重命中判断
智能调取详情
LLM 集成（如需要）

💡 最佳实践

文件命名：使用有意义的文件名
标签使用：为文件添加相关主题标签
时间标签：使用时间标签帮助时间轴匹配
批量操作：批量创建文件以提高效率
定期清理：删除不再需要的文件

⚠️ 注意事项

文件内容会进行向量嵌入，需要网络连接（首次）
LLM 功能需要配置 OpenAI API Key
Redis 可选，未配置时使用内存缓存
向量数据库存储在本地 vector_db 目录

🔄 数据存储

所有数据存储在：

~/.openclaw-autoclaw/workspace/context-manager/
├── agents/           # Agent 数据
├── vector_db/        # 向量数据库
├── cache/            # 缓存数据
└── logs/             # 日志文件

📈 性能优化

预期效果：

LLM 调用减少：60-80%
检索速度提升：10-50 倍
项目完成度显著提高

🔍 技术栈

向量数据库: ChromaDB
嵌入模型: Sentence Transformers
缓存: Redis + 内存
API: FastAPI
LLM: OpenAI GPT（可选）

🛠️ 故障排除

问题：向量数据库初始化失败

解决：删除 vector_db 目录后重试

问题：Redis 连接失败

解决：系统会自动使用内存缓存

问题：LLM 调用超时

解决：检查网络连接和 API Key 配置

📚 示例对话

用户：为 "研究项目" 创建一个 Agent
AI：✓ 已创建 Agent "research-project"

用户：将以下内容保存到研究项目的文件 "文献综述.md"：
深度学习是机器学习的一个子集...
AI：✓ 文件 "文献综述.md" 已创建并索引（索引时间: 452ms）

用户：在研究项目中搜索 "神经网络"
AI：找到 3 个相关文件：
1. 文献综述.md（相似度: 0.89）
2. 算法总结.md（相似度: 0.76）
3. 项目计划.md（相似度: 0.65）

用户：获取 "神经网络" 相关内容的详情
AI：已获取 2 个文件的详情。LLM 回复：
[详细答案]

🎓 更多信息

查看完整文档： cat ~/.openclaw-autoclaw/workspace/context-manager/README.md

运行示例： python ~/.openclaw-autoclaw/workspace/context-manager/example.py

🔗 相关技能

memory - 简单的内存管理
wecom-doc - 企业微信文档管理

Usage Guidance

What to consider before installing: - The bundle does not include the core 'context_manager' implementation it claims to use. skill.py expects to import that module from ~/.openclaw-autoclaw/workspace/context-manager; verify what code actually exists at that path before enabling the skill. - Because the skill adds that workspace path to Python's import path, any module named context_manager placed there will be executed by the skill. Only proceed if you trust the code in that directory or you control it. - The SKILL.md mentions optional use of an OpenAI API key and Redis. Only provide those secrets if you trust the skill and the code it imports; a malicious context_manager implementation could exfiltrate keys or data to remote endpoints. - The included demo (skill_demo.py) is safe and in-memory; consider running the demo first to verify behavior. Do not run the full skill until you inspect or install a trusted context_manager implementation and its dependencies (chromadb, sentence-transformers, etc.). - Recommended steps to reduce risk: inspect ~/.openclaw-autoclaw/workspace/context-manager for unexpected files, run the demo in an isolated environment, install third-party dependencies from trusted sources, and monitor network activity when first using the skill. If you can provide the missing context_manager implementation or confirm its source (e.g., a trusted GitHub repo), I can reassess the coherence and risk with higher confidence.

Capability Analysis

Type: OpenClaw Skill Name: autoglm-context-manager Version: 1.0.0 The skill bundle implements a legitimate context management system for AI agents, utilizing vector indexing (ChromaDB) and semantic search to manage long-term memory. The code in skill.py and skill_demo.py provides standard CRUD operations for agents and files, as well as search functionality. While the test scripts (final_test.py, simple_test.py) contain hardcoded local file paths (e.g., C:\Users\TIANXIN-01\...), these appear to be artifacts of a local development environment rather than malicious indicators. No evidence of data exfiltration, unauthorized execution, or prompt injection was found.

Capability Assessment

⚠ Purpose & Capability

The SKILL.md and description claim a full system (ChromaDB, sentence-transformers, Redis, optional OpenAI integration, FastAPI) but the distributed files do not include the core 'context_manager' implementation or vector DB code. Instead, skill.py appends ~/.openclaw-autoclaw/workspace/context-manager to sys.path and attempts to import a context_manager module that is not bundled. That means the skill delegates essential functionality to external code expected to live in the user's home workspace — a coherence gap between claimed capabilities and what's actually provided.

ℹ Instruction Scope

The SKILL.md instructs creating agents, saving files, and running example scripts under ~/.openclaw-autoclaw/workspace/context-manager/ and mentions network usage for embeddings/LLM. The included demo implementation (skill_demo.py) operates only in-memory and does not perform embeddings or network calls. Real embedding/LLM behavior would come from the absent context_manager module; SKILL.md also tells the user to run python example and cat README from that workspace, which implies the agent may read local files during normal operations.

⚠ Install Mechanism

There is no install spec and no packaged implementation of the heavy dependencies (ChromaDB, sentence-transformers, etc.). The skill expects external components to be installed in the user's workspace. The absence of a controlled install path means arbitrary or third-party code under ~/.openclaw-autoclaw/workspace/context-manager will be imported and executed when the skill runs.

✓ Credentials

No required environment variables or credentials are declared. SKILL.md documents optional OPENAI_API_KEY and Redis settings that are proportionate to the described (optional) LLM and cache features. Requesting an OpenAI key for optional LLM integration aligns with the stated functionality.

ℹ Persistence & Privilege

The skill does not request always:true and has no declared system-wide config access. However, it modifies sys.path to include a directory in the user's home workspace and will import whatever 'context_manager' exists there; that design means the skill will execute code from that workspace if present, which is a behavioral privilege to be aware of (not an automatic platform-level escalation, but still a runtime execution surface).

Version History

v1.0.0

Initial release: Long-term memory management for multi-agent systems with vector search and timeline filtering

Metadata

Slug autoglm-context-manager

Version 1.0.0

License MIT-0

All-time Installs 0

Active Installs 0

Total Versions 1

Frequently Asked Questions

What is Context Manager?

管理多个 AI Agent 的长期记忆与文件，通过向量和时间双重过滤，实现高效语义检索和减少 LLM 调用。 It is an AI Agent Skill for Claude Code / OpenClaw, with 93 downloads so far.

How do I install Context Manager?

Run "/install autoglm-context-manager" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Context Manager free?

Yes, Context Manager is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Context Manager support?

Context Manager is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Context Manager?

It is built and maintained by JanusSilence (@janussilence); the current version is v1.0.0.

More Skills

Context Manager