Memory MCP Server

Claude Code silently compresses older conversation turns to stay within its context window. You have seen it: "111 turns omitted." Every decision, every file path, every nuance from earlier in the session -- gone. This server captures live conversation snapshots every 5 prompts so that context is never lost.

Why This Exists

Claude Code's context compression is aggressive and invisible. Mid-conversation, it drops the exact turns where you discussed architecture, debugged a tricky issue, or agreed on a design. By the time you need that context, it is already gone.

Existing memory solutions bolt on basic RAG -- they index static documents and retrieve them on demand. That helps with long-term recall, but it does nothing for the conversation you are in right now.

This server solves both problems:

• Live mid-conversation capture. A hook fires every 5 prompts and writes a snapshot of recent turns to your vault. When Claude compresses old turns, the snapshots survive intact.
• Session-end summaries. When a session ends, a lightweight summary is written with wikilinks back to every snapshot, so you can trace the full conversation later.
• Semantic recall. On each new prompt, the most relevant memories from your vault are injected into context automatically. Past sessions inform current ones.

The result: Claude Code with a persistent, searchable memory that captures context as it happens, not after the fact.

Features

• Snapshot capture -- saves conversation turns to markdown every 5 prompts
• Session summaries -- end-of-session notes with topic, tools used, and links to snapshots
• Semantic search -- find related memories using Voyage AI embeddings and LanceDB
• Full-text search -- keyword search across all notes
• Obsidian-native -- writes standard markdown with frontmatter, works with any vault
• Auto-recall -- injects relevant memories into context on every prompt
• File watcher -- auto-reindexes when vault files change on disk

Quick Start

git clone https://github.com/MaybeLOL/memory-mcp-server.git
cd memory-mcp-server
npm install && npm run build

Then add the server to ~/.claude/mcp.json and the hooks to ~/.claude/settings.json (see Configuration below).

How It's Different

Most memory tools treat recall as a lookup problem. This server treats it as a continuity problem -- making sure context from 50 turns ago is still available even after Claude Code compresses it away.

Configuration

1. MCP Server

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["/path/to/memory-mcp-server/dist/index.js"],
      "env": {
        "VAULT_PATH": "/path/to/your/vault",
        "DB_PATH": "/path/to/lancedb-data",
        "VOYAGE_API_KEY": "your-voyage-api-key"
      }
    }
  }
}

DB_PATH is where LanceDB stores its vector index. The directory is created automatically.

2. Hooks

Add these hooks to ~/.claude/settings.json. If you already have a settings.json, merge the hooks section into your existing config.

macOS / Linux / Git Bash on Windows:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "VAULT_PATH=\"/path/to/vault\" VOYAGE_API_KEY=\"your-key\" node \"/path/to/memory-mcp-server/dist/recall.js\"",
            "timeout": 15
          },
          {
            "type": "command",
            "command": "VAULT_PATH=\"/path/to/vault\" node \"/path/to/memory-mcp-server/dist/mid-capture.js\"",
            "timeout": 15
          }
        ]
      }
    ],
    "SessionEnd": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "VAULT_PATH=\"/path/to/vault\" node \"/path/to/memory-mcp-server/dist/capture.js\"",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Windows (cmd/PowerShell): The inline VAR=value command syntax does not work. Use wrapper scripts or set environment variables in your system settings. If Claude Code uses Git Bash as its shell, the above syntax works as-is.

Replace /path/to/vault and /path/to/memory-mcp-server with your actual paths.

3. Environment Variables

MCP Tools

How It Works

Mid-Conversation Capture

The mid-capture hook runs on every UserPromptSubmit. It maintains a counter at ~/.claude/session-turn-count.json and every 5th prompt:

1. Reads the conversation transcript JSONL
2. Extracts all turns since the last snapshot
3. Writes them to <vault>/conversations/YYYY-MM-DD-session-<slug>-<id>-snap-N.md

Non-capture prompts exit in under 500ms (just a counter increment).

Session-End Summary

When a session ends, the capture hook writes a lightweight summary containing:

• Topic, tools used, and working directory
• First and last turn for context
• Wikilinks to all mid-session snapshots

Recall

On each prompt, the recall hook embeds the user's message via Voyage AI and compares it against all vault notes, then injects the top semantic matches as a system message.

Note: The recall hook re-embeds all notes on every prompt. For large vaults (500+ notes), this may be slow or exceed the 15s timeout. Consider increasing the timeout or reducing vault size if needed.

Prerequisites

• Node.js 20+
• A Voyage AI API key (for embeddings)
• An Obsidian vault or any folder of markdown files

Troubleshooting

MCP server not connecting: Check that VAULT_PATH, DB_PATH, and VOYAGE_API_KEY are all set. Run manually to see errors:

VAULT_PATH=/path/to/vault DB_PATH=/path/to/db VOYAGE_API_KEY=your-key node dist/index.js

Hooks not firing: Verify your ~/.claude/settings.json is valid JSON. Check that file paths in hook commands are correct and absolute.

`npm install` fails on LanceDB: The @lancedb/lancedb package includes native binaries. If you are behind a proxy or on an unsupported platform, the download may fail. See LanceDB docs for platform support.

Recall is slow: For vaults with many notes, the recall hook may time out. Increase the timeout value in your hook config, or reduce the number of notes in your vault.

License

MIT