AgentRecall-MCP

概述

AgentRecall 是一个 MCP 服务器，通过日记录制、意识复合和语义召回功能为 AI 代理提供持久记忆系统。它实现思考-执行-反思的质量循环，使 AI 代理能够跨会话学习。系统对记忆进行索引，应用权重和衰减算法，并通过反馈循环实现持续改进。它可以与现有 Git 仓库、Claude 记忆系统及其他系统协作，通过引导导入功能无缝集成。

试试问 AI

装完之后，这里有 5 个你可以让 AI 做的事：

• arstatusarstatus()

  Shows status board across all projects, showing pending work, blockers, relevance scores, and recommended next actions

  试试:Show me my current projects and what I should work on next试试:What projects are blocked and why?试试:Display my project status board
• arsavearsave()

  Saves the current session by writing journal entries, consolidating insights, and updating palace memory

  试试:Save my work from this session试试:End the session and save all my insights试试:Compound my learning from today's work
• arstartarstart(project: string)

  Loads deep context for a specific project including palace rooms, corrections, and task-specific recall

  试试:Start working on my project 'novada-mcp'试试:Load context for my 'agentrecall' project试试:Begin session for project 'aam'
• arsaveallarsaveall()

  Batch saves all parallel sessions across projects, scanning, merging, deduplicating insights, and updating awareness in one operation

  试试:Save all my parallel sessions from today试试:Batch save all open projects试试:Consolidate and save all my work across sessions
• arbootstraparbootstrap(options?: { dryRun?: boolean, project?: string })

  Scans the machine for existing projects and imports them automatically, supporting migration from other memory systems

  试试:Import my existing projects from other tools试试:Scan my machine for projects to import试试:Bootstrap from my Claude AutoMemory
• bootstrap_scanbootstrap_scan()

  Discovers existing projects and context on the machine without importing them

  试试:What projects would be imported if I bootstrap?试试:Scan for existing projects without importing试试:Show me what's ready to be imported
• bootstrap_importbootstrap_import(scan_result: any)

  Imports selected projects from a previous scan result

  试试:Import these specific projects from my scan试试:Bring in my 'agentrecall' and 'novada-mcp' projects试试:Import just these three projects
• arcorrectionarcorrection(correction: string)

  Captures and stores corrections to improve agent behavior in future sessions

  试试:Remember to use Playwright instead of bb-browser试试:Note that I prefer light themes over dark ones试试:Capture this as a correction for future sessions
• arinsightarinsight(insight: string)

  Adds insights to the knowledge base that can be recalled across sessions and projects

  试试:Add this insight about server-rendered cards to my knowledge试试:Remember that stemming improves keyword recall试试:Capture this as a reusable insight
• arprefetcharprefetch(query: string)

  Pre-fetches relevant memory before starting a new session or task

  试试:Find relevant memories about deploying to Vercel试试:Prefetch context about multi-agent workflows试试:Retrieve insights about async API patterns

可对比工具

安装

npm install -g agent-recall-mcp

{
  "mcpServers": {
    "agent-recall": {
      "command": "npx",
      "args": ["agent-recall-mcp"]
    }
  }
}

FAQ

AgentRecall 的核心命令是什么？

两个基本命令是每个会话开始时的 `/arstatus`（选择项目）和结束时的 `/arsave`（保存记忆并复合见解）。其他命令包括 `/arstart`（加载上下文）、`/arsaveall`（批量保存）和 `/arbootstrap`（导入现有记忆系统）。

AgentRecall 是否与其他记忆系统兼容？

是的，AgentRecall 可以通过 `/arbootstrap` 命令自动发现并导入 Claude AutoMemory、Git 仓库、CLAUDE.md 文件等系统的记忆。它会扫描您的机器并导入现有上下文，不会覆盖源文件。

AgentRecall-MCP 对比