rlm-claude

概述

RLM 通过实现一个 MCP 服务器解决 Claude Code 的上下文窗口限制问题，在压缩操作前自动保存对话历史。它实现了两种内存系统：用于保存关键决策和事实的 Insights，以及用于完整对话片段的 Chunks。该服务器提供高级功能，包括结合 BM25 和余弦相似性的语义搜索、模糊搜索、多项目支持、智能保留策略和子代理分析功能。RLM 挂钩到 Claude Code 的事件中，为跨会话提供零配置的持久化内存。

试试问 AI

装完之后，这里有 5 个你可以让 AI 做的事：

• rlm_remember

  Save decisions, facts, preferences with categories and importance levels

  试试:Remember that the client prefers 500ml bottles试试:Save the decision to use PostgreSQL for the backend试试:Remember that we should always deploy LOCAL → VPS
• rlm_recall

  Search insights by keyword, category, or importance

  试试:What did we decide about the API architecture?试试:Show me critical insights about deployment试试:Find facts about authentication systems
• rlm_forget

  Remove an insight from memory

  试试:Forget the temporary note about testing credentials试试:Remove the old pricing decision from insights
• rlm_status

  Get system overview including insight count and chunk stats

  试试:What's the current status of my memory?试试:Show me memory statistics
• rlm_chunk

  Save conversation segments with typed categorization

  试试:Save this discussion about API redesign as a session chunk试试:Take a snapshot of our current conversation试试:Save this debug information for later reference
• rlm_peek

  Read a chunk in full or partially by line range

  试试:Show me the first 50 lines of chunk 2026-01-18_MyProject_001试试:Read the full conversation chunk from yesterday
• rlm_grep

  Regex search across all chunks with optional fuzzy matching

  试试:Search for 'authentication' across all chunks试试:Find mentions of 'bug #42' with fuzzy matching
• rlm_search

  Hybrid search using BM25 + semantic similarity across chunks and insights

  试试:Search for information about our API architecture decisions试试:Find content related to deployment strategies
• rlm_list_chunks

  List all chunks with metadata

  试试:Show me all my saved conversation chunks试试:List all session chunks from this week
• rlm_sessions

  Browse sessions by project or domain

  试试:Show me all sessions for MyApp project试试:List sessions from the backend domain
• rlm_retention_preview

  Preview what would be archived in a dry-run

  试试:What chunks would be archived if I ran retention now?试试:Preview the purge operation
• rlm_retention_run

  Archive old unused chunks and purge ancient ones

  试试:Run the retention process to clean up old chunks试试:Archive unused chunks and purge anything older than 6 months

可对比工具

安装

pip install mcp-rlm-server[all]

uv tool install mcp-rlm-server[all] --python 3.12

{
  "mcpServers": {
    "rlm": {
      "command": "python",
      "args": ["-m", "mcp_server.server"]
    }
  }
}

FAQ

RLM 如何在会话间持久化内存？

RLM 挂钩到 Claude Code 的 /compact 事件，在上下文丢失前自动保存快照。内存存储在 ~/.claude/rlm/ 中，使用 JSON 格式存储 Insights，使用文件格式存储对话历史片段。

Insights 和 Chunks 有什么区别？

Insights 存储关键决策、事实和偏好，带有类别和重要性级别。Chunks 存储完整的对话片段，类型分为快照、会话或调试。Insights 按重要性加载，Chunks 则按需搜索和检索。

rlm-claude 对比