Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

Viking Memory System Ultra

v1.4.0

Viking Memory System Ultra - 分层记忆基座。特性:动态回流(语义晋升)、智能权重(对数增长)、可逆归档(多粒度摘要)。核心脚本:sv_write, sv_read, sv_find, sv_autoload, sv_promote, sv_weight, sv_compress。

0· 120·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for tsangho/viking-memory-ultra.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Viking Memory System Ultra" (tsangho/viking-memory-ultra) from ClawHub.
Skill page: https://clawhub.ai/tsangho/viking-memory-ultra
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install viking-memory-ultra

ClawHub CLI

Package manager switcher

npx clawhub@latest install viking-memory-ultra
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Suspicious
medium confidence
!
Purpose & Capability
The skill's purpose (local layered memory manager) matches the instructions that manipulate files under agent/memories/. However the SKILL.md repeatedly references many 'core scripts' (sv_write, sv_read, sv_promote, sv_compress_v2.sh, sv_archive_summary.sh, sv_decompress.sh, sv_review, sv_list, etc.) but the package contains no binaries or install spec. That mismatch (claims scripts but provides nothing and requests no dependencies) is an incoherence: either the runtime environment must already provide these tools, or the skill is incomplete.
!
Instruction Scope
Instructions tell agents to read, write, compress, promote and scan cold/archive layers for semantic similarity and to sync 'important' memories to viking-global/shared/memory/hot/. All of this stays within the stated memory-management domain, but the 'sync to shared space' step implies sharing potentially sensitive agent data with other agents or a shared location. SKILL.md does not define the shared endpoint, access controls, or required credentials, so the data-exposure behavior is underspecified and potentially risky.
Install Mechanism
No install spec is provided (instruction-only), which is lower risk in terms of arbitrary code downloads. However, because the skill expects numerous shell scripts and scheduled operations, the absence of install instructions or bundled code creates an operational gap and potential for error or misconfiguration if those scripts are missing or come from an untrusted source.
Credentials
The skill declares no required environment variables or credentials, which is good for limiting secret exposure. Yet it instructs syncing to a shared global path (viking-global/shared/...), scanning archives, and possibly interacting with other agent skills (daily-digest). Those actions often require filesystem permissions, mounts, or network/configured shared storage—none of which are documented or requested. The lack of declared credentials or config for shared storage is suspiciously under-specified.
Persistence & Privilege
always:false and normal autonomous invocation are used. The skill does not request permanent platform-wide presence and does not instruct modifying other skills or system-wide agent configs. Scheduling (weekly compression at 2:00) is documented but there is no instruction to programmatically modify cron or system scheduler in the SKILL.md.
What to consider before installing
This skill appears to implement a local layered-memory system, but proceed carefully. Key points to check before installing or enabling: 1) Source verification: the registry Owner ID shown in metadata does not match the _meta.json ownerId—ask the publisher to explain and provide a verified source or homepage. 2) Missing scripts: SKILL.md references many shell scripts but none are bundled or installed—ask for an install guide or the actual script sources; do not let untrusted scripts be fetched automatically. 3) Shared sync risk: 'sv_write_v2 --hot' is documented to copy 'important' memories to viking-global/shared/...; understand who has access to that shared location and whether that will leak private data. 4) Run first in an isolated/test agent or sandbox and audit the actual scripts and any cron/scheduling changes. 5) If you must use it in production, require explicit configuration for the shared endpoint and permissions, add logging/approval for sync operations, and ensure backups and access controls are in place. If the publisher cannot clarify the owner mismatch and supply the missing script sources or an install procedure, treat the skill as untrusted.

Like a lobster shell, security has layers — review code before you run it.

latestvk97f3m475d7cvemwdz934fjcdh849jft
120downloads
0stars
3versions
Updated 3w ago
v1.4.0
MIT-0

Viking 记忆系统 - SKILL.md

版本:v1.4(Phase 4 复习状态追踪) 激活日期:2026-03-27 更新日期:2026-03-29 这是 Viking 记忆系统的核心使用规范,所有 Agent 必须遵循


🏠 系统概述

Viking 是我们的分层记忆基座,解决"记忆太多无法高效检索"的根本矛盾。

热记忆(hot) → 温记忆(warm) → 冷记忆(cold) → 归档(archive)
   0-7天         8-29天        30-89天       90天+

📁 目录结构

每个 Agent 的 Viking 记忆位于 $SV_WORKSPACE/agent/memories/

agent/memories/
├── daily/          # 每日工作笔记(由 daily-digest skill 读取)
├── hot/            # L0: 完整记忆细节(0-7天)
├── warm/           # L1: 压缩关键要点(8-29天)
├── cold/           # L2/L3: 极简标签/索引(30-89天)
├── archive/        # L4: 长期归档(90天+,可搜索触发回忆)
├── long-term/      # 永久知识沉淀
├── meetings/       # 会议记录
└── config/         # 层级配置

🔧 核心命令

写入记忆(sv_write)

# 基本写入 → hot 层
sv_write "agent/memories/hot/2026-03-27-重要发现.md" "# 内容"

# 指定层级
sv_write "agent/memories/warm/2026-03-27-周报摘要.md" "# 内容"

# 重要记忆(标记 important=true,跳过自动压缩)
sv_write "agent/memories/hot/2026-03-27-战略决策.md" "# 内容" --importance high

读取记忆(sv_read)

sv_read "agent/memories/hot/2026-03-27-重要发现.md"

同步到共享空间(重要记忆)

# 重要记忆同步到 viking-global/shared/memory/hot/,让其他 Agent 可见
sv_write_v2 --hot "agent/memories/daily/2026-03-27.md" "# 重要内容"

📝 Frontmatter 规范(必须遵循)

所有写入 Viking 的记忆必须包含标准 frontmatter:

---
id: mem_20260327_153000
title: "今日重要决策"
importance: high        # high | medium | low | important(important=永久保留)
important: false        # true = 永不压缩
tags: [决策, 战略, 猫经理]
created: 2026-03-27T15:30:00+08:00
last_access: 2026-03-27T15:30:00+08:00
access_count: 1
context_correlation: 1.0  # Phase 2: 上下文相关性系数 (0.5-1.5)
retention: 90           # 保留天数(重要/important 可设更长)
current_layer: L0       # L0 | L1 | L2 | L3 | L4
target_layer: hot       # hot | warm | cold | archive
weight: 10.0           # Phase 2 权重 = factor × decay × ln(count+1) × corr
# 复习状态追踪(Phase 4)
review_status: pending_review  # pending_review | reviewing | reviewed | mastered
last_reviewed: null
next_review: null
review_count: 0
---

# 今日重要决策

## 决策内容...

⚙️ 自动压缩规则

文件年龄自动迁移
0-7天hot/ (L0)
8-29天warm/ (L1)
30-89天cold/ (L2/L3)
90天+archive/ (L4)

重要标记important: trueimportance: important)= 永不自动压缩


🚨 使用规范

必须遵循

  1. 重要决策 → 写入 agent/memories/hot/,标记 importance: important,同步到共享空间
  2. 每日工作 → 写入 agent/memories/daily/YYYY-MM-DD.md(与 daily-digest skill 对接)
  3. 学到新知识 → 写入 agent/memories/hot/learning-YYYY-MM-DD.md
  4. 会议结论 → 写入 agent/memories/meetings/YYYY-MM-DD.md
  5. 团队协调 → 重要内容用 sv_write_v2 --hot 同步到 viking-global

禁止

  1. ❌ 不要把所有东西都塞 hot/(会导致 hot 层膨胀)
  2. ❌ 不要跳过 frontmatter(否则压缩脚本无法处理)
  3. ❌ 不要在 hot/ 存超过7天的无关内容
  4. ❌ 不要在 agent/memories/ 外的地方建立记忆体系(memory/ 是平面的,Viking 是分层的)

🔄 压缩调度

  • 每周日凌晨 2:00sv_compress_v2.sh --force 自动执行
  • 手动触发sv_compress_v2.sh --dry-run 预览变更

📦 Phase 4: 复习状态追踪 (v1.4)

解决"用户复习了哪些内容、如何判定复习状态"的跟踪问题

复习状态定义

状态说明触发条件
pending_review待复习新建记忆的默认状态
reviewing复习中用户开始复习
reviewed已复习完成一次复习,自动计算下次复习日期
mastered已掌握多次复习后标记,自动移动到 archive 层

艾宾浩斯复习曲线

复习完成后自动计算下次复习日期(简化版):

复习次数间隔
第1次1天后
第2次3天后
第3次7天后
第4次15天后
第5次+30天后

复习状态管理命令

# 标记复习状态
sv_review mark <file> <status>

# 列出指定状态的记忆
sv_review list pending_review
sv_review list reviewed

# 显示统计
sv_review stats

# 显示今日需复习的记忆
sv_review due

示例

# 标记为已复习(自动更新 last_reviewed 和 next_review)
sv_review mark "agent/memories/hot/2026-03-29.md" reviewed

# 标记为已掌握(自动移动到 archive 层)
sv_review mark "agent/memories/hot/重要概念.md" mastered

# 查看统计
sv_review stats

sv_list 支持复习状态过滤

# 只显示待复习的记忆
sv_list --review-status pending_review

# 只显示已复习的记忆
sv_list --review-status reviewed

📦 Phase 3: Archive 可逆性 (v1.3)

归档不再是"彻底遗忘",而是多粒度压缩存储,支持按需解压恢复

核心改进

旧方案新方案
归档 = 丢弃正文,只留标题归档 = 多粒度摘要 + 完整内容存档
需要时无法找回细节可随时解压完整内容
只能线性遗忘支持选择性回忆

新增脚本

脚本功能
sv_archive_summary.sh生成多粒度摘要,保留完整内容
sv_decompress.sh按需解压,恢复到目标层

使用方式

# 归档时自动生成摘要(集成到 sv_compress_v2.sh)
sv_compress_v2.sh --force

# 手动为 archive 文件生成摘要
sv_archive_summary.sh <file> --keep

# 查看摘要
sv_decompress.sh <file>

# 显示完整内容
sv_decompress.sh <file> --show

# 恢复到 hot 层
sv_decompress.sh <file> --restore

frontmatter 新增字段

summary: "一句话摘要 | 段落摘要"
full_content_file: "filename.md.archive.full"

⚖️ Phase 2: 权重公式优化 (v1.2)

改进 access_count 增长模式,引入上下文相关性系数

旧公式 vs 新公式

项目旧公式新公式
访问次数增长线性 (count+1)对数 ln(count+1)
上下文相关性❌ 无✅ 0.5-1.5 系数
防止记忆霸权❌ 无法防止✅ 旧记忆不会无限膨胀

新公式

W = importance_factor × (1/(days+1)^0.3) × ln(access_count+1) × context_correlation

参数说明

参数说明范围
importance_factor重要性因子high=3.0, medium=1.5, low=0.5
time_decay时间衰减0-1,越老越小
ln(access_count+1)对数增长防止无限膨胀
context_correlation上下文相关性0.5-1.5(默认1.0)

context_correlation 设置建议

场景系数说明
高相关任务访问1.3-1.5当前工作直接相关
正常访问1.0例行加载
低相关但被召回0.7-0.9动态回流晋升

使用方式

# 单独计算
sv_weight.sh <file>

# 更新权重(默认 context_correlation=1.0)
sv_weight.sh <file> --update

# 指定上下文相关性
sv_weight.sh <file> --update --context-correlation 1.5

# 集成到加载流程
sv_autoload.sh --update-weight

🔥 动态回流机制 (v1.1 新增)

当检测到当前任务与 Cold/Archive 记忆语义相似时,自动晋升至 Hot 层

背景

传统系统的痛点:记忆只能向下迁移(hot→warm→cold→archive),永远不会回来。 这导致"三个月前做过类似项目,但系统已经忘了"的问题。

解决方案

动态回流机制:每次加载记忆时,自动扫描 cold/archive 层,找出语义相关的记忆并晋升。

核心脚本

# 独立运行
sv_promote.sh --context "当前任务描述"
sv_promote.sh --dry-run  # 预览模式
sv_promote.sh --threshold 0.7  # 调整阈值

# 集成到加载流程
sv_autoload.sh --promote

晋升流程

  1. 语义相似度判断:使用 LLM 判断当前上下文与历史记忆的相关性
  2. 阈值过滤:相似度 ≥ 0.7 才晋升(可调整)
  3. 重要记忆保护importance: high/important 的记忆不受自动晋升影响
  4. 原子性操作:先写临时文件,再移动,避免数据丢失

技术细节

参数默认值说明
--threshold0.7相似度阈值,0.0-1.0
--layercold,archive扫描的源层级
--context自动获取当前任务上下文

API 支持

  • 优先使用 NVIDIA Qwen(如已配置)
  • Fallback: MiniMax M2.5

注意事项

  • ⚠️ 晋升是单向的(cold/archive → hot),不会降级
  • ⚠️ 重要记忆(high/important)不会被自动晋升,保持原层级
  • ⚠️ 每次最多处理 50 条记忆,防止 API 过载
  • 💡 建议在会话开始时用 sv_autoload.sh --promote,而非频繁独立调用

💡 实用技巧

如何判断写哪个层级:

  • 有细节、有过程、有价值 → hot/L0
  • 只有结论、要点 → warm/L1
  • 只是一个标签/关键词 → cold/L2

团队共享记忆:

  • 猫经理的重要决策 → 同步到 viking-global/shared/memory/hot/
  • 各 Agent 每天早上应读取共享 hot 层:sv_read 重要共享文件

📂 相关文件

  • 脚本:$VIKING_HOME/scripts/sv_*.sh
  • 配置:$VIKING_HOME/config.yaml
  • 共享空间:$VIKING_GLOBAL/shared/memory/

Comments

Loading comments...