Viking 记忆系统

支持五级自动记忆管理和重要记忆保护，集成飞书群聊会话自动保存与向量相似度检索，提升跨 Agent 记忆共享。

Security Scan

功能与文件总体一致，但包声明与运行时脚本不匹配——脚本访问本地会话文件、调用本地/远端 LLM 服务并依赖未声明环境变量/插件，需在生产环境安装前人工审查配置与端点。

Purpose & Capability

技能名/描述为记忆管理与飞书集成，代码确实实现了分层存储、自动保存、向量检索与降级逻辑 — 与目的匹配。但 registry metadata declares no required env or binaries while scripts expect and use several external services/tools (OPENCLAW_BIN, OLLAMA_HOST, EMBED_MODEL, feishu plugin, sv CLI). The omission of these runtime dependencies in the skill metadata is an incoherence that could hide important configuration or network endpoints.

Instruction Scope

SKILL.md instructs adding cron and enabling feishu plugin; scripts read local agent session files (~/.openclaw/agents/.../sessions/*.jsonl), run OpenClaw CLI to call agents/LLMs, and call an embedding service (curl to OLLAMA_HOST). These actions will collect chat/session content and send it to LLM/embedding endpoints (local IP default or whatever OLLAMA_HOST is set to). The instructions do not enumerate these data flows or the required host/credentials explicitly, giving the agent broad discretion to process and transmit user conversations.

Install Mechanism

No automated install spec — user copies files and sets up cron as described. That means nothing arbitrary is fetched/installed by the skill itself (lower install risk). The code bundle is provided in the skill, so runtime behavior depends on local execution under user control.

Credentials

Metadata lists no required env vars, but multiple scripts read/use environment variables and external endpoints: OLLAMA_HOST (default http://192.168.5.110:11434), OPENCLAW_BIN, EMBED_MODEL, AGENT_NAME, etc. The skill also assumes the feishu plugin (which will have its own credentials) and sv/sv build-index command. Requesting access to agent session files and calling external embedding/LLM services is proportionate to memory/search features, but the skill fails to declare these requirements and does not make endpoint/credential usage explicit — this increases audit burden and risk of unintended data exposure.

Persistence & Privilege

always:false and user-invocable:true — no forced always-on privilege. The skill writes/reads files under ~/.openclaw/viking-* and may be scheduled by user-configured cron jobs (SKILL.md asks user to add cron). That persistent scheduling is under user control, but once enabled the cron and session hooks will process chat logs automatically. The skill does not appear to modify other skills' configuration, but it does register tools and read openclaw skills list if auto-record enabled.

What to consider before installing

What to check before installing: - Do not enable cron or the feishu automatic hooks until you review and test everything locally. - Inspect and explicitly set OLLAMA_HOST, OPENCLAW_BIN, EMBED_MODEL, AGENT_NAME in a safe test environment. The default OLLAMA_HOST points to a LAN IP (192.168.5.110) — confirm who controls that host. Any embedding/LLM host will receive your session data. - Confirm the feishu plugin installation and its credentials; ensure you understand where chat data flows and who can read it. - Review the scripts that read ~/.openclaw/agents/.../sessions/*.jsonl and memory workspaces to ensure they only process data you consent to store in Viking workspaces. - Run the scripts in a non-production agent first, with mocked or sanitized session files, and verify where network calls go (curl calls and openclaw agent invocations). - If you only want local operation, ensure the embedding/LLM endpoints are local and isolated; otherwise consider disabling embedding/network calls. - Consider adding explicit declarations for required env vars and endpoints in the skill metadata and restrict cron to an account with limited access. If you want, I can produce a short checklist and recommended config values to test this skill safely.

Viking 记忆系统技能

功能

• 五级记忆层级自动管理（L0→L4）
• 重要记忆保护（永不降级）
• 向量相似度提及检测（动态重要性）
• 飞书群聊会话自动保存（需飞书插件）
• 跨 Agent 共享（global workspace）
• 自动降级与遗忘算法（LLM 压缩）

使用方式

1. 安装技能

确保以下文件已复制到你的技能目录：

cat-viking-memory/
├── config.json
├── SKILL.md
├── memory-pipeline/     # CLI 主入口和脚本
├── simple-viking/       # 向量搜索
└── references/          # 文档
    ├── README.md        # 详细使用说明
    └── 飞书集成说明.md

2. 基本配置

编辑 config.json 中的配置：

{
  "memory_workspace": "~/.openclaw/viking-{agent}",
  "auto_save_enabled": true,
  "feishu_integration": {
    "enabled": false
  }
}

3. 定时任务（推荐）

每天凌晨 3 点自动执行记忆降级：

# crontab -e
0 3 * * ~/.openclaw/workspace/agents/maozhuli/cat-viking-memory/memory-pipeline/memory-tier-cron.sh

4. 飞书自动保存（可选）

如需飞书群聊会话自动保存：

1. 确保 feishu 插件已安装并启用
2. 修改 config.json：

   "feishu_integration": {
     "enabled": true,
     "check_interval_ms": 300000,
     "timeout_ms": 1800000
   }
   

3. 在 feishu 插件配置中启用 SessionManager
4. 参考 references/飞书集成说明.md 详细步骤

5. 常用命令

# 加载记忆上下文
memory-pipeline mp_autoload

# 保存记忆
memory-pipeline mp_global "任务完成说明"

# 搜索记忆
sv_find "关键词"

# 手动触发降级（测试）
memory-pipeline/memory-tier-cron.sh

相关文档

• 详细使用说明：references/README.md
• 飞书集成指南：references/飞书集成说明.md
• 记忆文件格式：见 references/README.md 的"记忆文件格式"章节

技术架构

存储流程：
输入 → Ontology映射 → Viking存储 → 向量索引 → 检索

层级管理：
L0(0-1天) → L1(2-7天) → L2(8-30天) → L3(30-90天) → L4(90天+)

自动保存：
会话结束 → memory-session-hook.sh → mp_global → 存储

飞书集成：
FeishuSessionManager → 检测超时 → 触发 Hook → 保存

注意事项

• 存储规范：必须使用 memory-pipeline 命令存取，不要直接操作文件
• 重要记忆：使用 --important 标记永不忘记的关键信息
• 定期检查：建议每周检查 crontab 是否正常执行
• 权限：确保脚本有执行权限 chmod +x memory-pipeline/*.sh

版本历史

• v1.0.0 - 初始版本
  • 五级记忆层级完整实现
  • 飞书会话自动保存集成
  • 向量搜索与语义检索
  • 自动降级与遗忘算法

---

基于 OpenViking 思想设计，专为 OpenClaw Agents 打造