360 Secure SkillHub
360 安全技能中心

vestige

该技能封装 vestige-mcp 4.2.0 本地 MCP 服务，为 AI Agent 提供链式决策守卫与记忆：smart_ingest 写入带收据的记忆，recall 按精确 id 或标签检索，causal_walk 与 forgotten_lesson 从失败记忆回溯因果，receipt 校验并重放写入，codebase 记录仓库变更；另附需 jq 的 vmem 脚本。

samvallad33
samvallad33
热度 60.47
最近更新 2026-10-16 07:16:32
分类：知识依赖：MCP 服务 / 代码执行应用场景：笔记整理 / 研究 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add samvallad33/vestige
Metadata
name
vestige
description
Vestige 4.2.0, an agent-agnostic guard for your AI agent. It judges the whole chain of steps, stops it before it does damage, keeps a hash-chained receipt of every decision, and saves that chain to memory so an old yes is never permission to act now. Local Rust MCP server, exact-handle recall, causal_walk from a failure back to its cause.

Vestige Skill

Vestige is an agent-agnostic guard for your AI agent. It judges the whole chain of steps, not just one command, and stops it before it does damage, keeping a receipt of every decision. It works with Claude Code, OpenClaw, any agent, and it saves that chain to memory so an old yes is never treated as permission to act now.

This skill wraps the official vestige-mcp server, version 4.2.0. Memory lives on Strata, an append-only hash-chained log: every write returns a receipt, and nothing is found by similarity. A memory is found by its exact id or an exact tag.

This skill is not OpenClaw default memory and not a plugins.slots.memory plugin. OpenClaw's memory-core stays as it is; Vestige runs beside it as an MCP server.

Install

npm install -g vestige-mcp-server@4.2.0
which vestige-mcp          # macOS / Linux
where vestige-mcp          # Windows

Put the absolute path in OpenClaw's MCP config (GUIs do not inherit your shell PATH or expand ~):

{
  "mcp": {
    "servers": {
      "vestige": {
        "command": "<absolute path from which vestige-mcp>",
        "args": []
      }
    }
  }
}

For a per-project store add "args": ["--data-dir", "<absolute dir>"].

Tools to use

Tool Use
session_start Open intentions, status and code memories at the start of a session
smart_ingest Save a memory (content, node_type, tags); returns a receipt
recall Find by exact handle: a full memory id or an exact, case-sensitive tag. Free text returns similarity_disabled
causal_walk Walk back from a failure along recorded causal edges: start_points=[{kind:"logged_write", node_id:"<failure memory id>"}]. Results are hypotheses, not proven causes
forgotten_lesson From a failure memory, find earlier lessons on recorded edges that are fading
receipt get shows what a write did; replay re-derives it from the log
codebase ingest_repo records a checkout's commits, reverts, cherry-picks and Fixes: trailers as change records; remember_decision / remember_pattern anchor decisions to files
selftest Planted-cause selftest on a throwaway log

Tag every memory on purpose: tags are the index.

When to use

  • A decision, correction or verified fact worth keeping: smart_ingest with a narrow topic tag
  • "What did we decide about X?": recall with that tag
  • "Why did this break?": save the failure as an event, then causal_walk and forgotten_lesson from its id
  • "Prove that write": receipt with the memory id

Never save API keys, passwords or other secrets.

Helper

vmem ships with this listing and resolves vestige-mcp with command -v. It needs jq.

vmem recall <tag-or-id>            # exact handle
vmem save "text" [tag]             # smart_ingest, optional tag
vmem walk <failure-memory-id>      # causal_walk from a logged write
vmem receipt <memory-id>           # receipt get
vmem selftest

Data location

Default data directory (override with --data-dir):

  • macOS: ~/Library/Application Support/com.vestige.core/
  • Linux: ~/.local/share/vestige/core/
  • Windows: %APPDATA%\vestige\core\

Credits