How to Add Persistent Memory to Claude Code in 5 Minutes

The Problem: Claude Code Forgets Everything

You spend an hour explaining your project architecture to Claude Code. You walk through your database schema, your naming conventions, the reason you chose that specific folder structure. Claude gives great answers. Then you close the terminal.

Next session? Gone.Every conversation, every decision, every piece of context — wiped clean. Claude Code has no persistent memory. You start over from scratch, every single time.

This is not a minor inconvenience. For teams that use Claude Code daily, the repeated context-setting wastes hours each week and leads to inconsistent outputs. Your AI assistant should remember what you told it yesterday.

The Solution: MemPalace MCP Server

MemPalace is an open-source memory system that gives Claude Code persistent memory through the Model Context Protocol (MCP). It stores every conversation verbatim, generates vector embeddings for semantic search, and retrieves the right context automatically when Claude needs it. Everything runs locally on your machine — no cloud, no API keys, no cost.

Prerequisites

• ✓Python 3.9+ — run python --version to check
• ✓Claude Code CLI — installed and authenticated

One command installs MemPalace and all its dependencies — ChromaDB for vector storage, SQLite for metadata, and Sentence Transformers for local embeddings.

pip install mempalace

First-time install downloads the all-MiniLM-L6-v2 embedding model (~80MB). This only happens once.

Create the local memory directory structure. This sets up the palace with its Wings, Rooms, and Halls — the organizational hierarchy for your memories.

mempalace init

You should see output like:

🏛️  Initializing MemPalace...
   Created: ~/.mempalace/
   Created: ~/.mempalace/palace.db
   Created: ~/.mempalace/vectors/
   ✓ Palace initialized with default Wing: "main"
   ✓ Room "general" created in Wing "main"
   ✓ MCP server ready on stdio

Your palace is ready. Connect it to an MCP client to start remembering.

Add the MemPalace MCP server to your Claude Code configuration. Open your Claude Code settings file and add the following to the mcpServers section:

{
  "mcpServers": {
    "mempalace": {
      "command": "mempalace",
      "args": ["serve", "--mcp"],
      "env": {}
    }
  }
}

Where to add this: Your Claude Code settings file is typically at ~/.claude/settings.json. If the file already has other MCP servers configured, just add the "mempalace" entry inside the existing mcpServers object.

Alternatively, you can run the shortcut command to do this automatically:

mempalace connect claude-code

Start a new Claude Code session and test the memory integration. Try this conversation:

You: Remember that our API uses snake_case for all endpoints
     and we deploy to AWS us-east-1.

Claude: I've stored that in memory. Your API uses snake_case
        endpoints and deploys to AWS us-east-1.

--- close terminal, open a new session ---

You: What conventions does our API use?

Claude: Based on my memory, your API uses snake_case for all
        endpoints and you deploy to AWS us-east-1.

That is it. Claude Code now has persistent memorythat survives across sessions. Every decision you share, every context you provide — it stays.

What Happens Behind the Scenes

When Claude Code stores a memory through MemPalace, four things happen in sequence:

1. 1

   Verbatim Storage

   The full conversation text is saved to SQLite with metadata (timestamps, tags, room). Nothing is summarized or lost.

2. 2

   Vector Embeddings

   Sentence Transformers generates a 384-dimensional vector embedding of the content using the all-MiniLM-L6-v2 model, running entirely on your CPU.

3. 3

   Semantic Indexing

   ChromaDB indexes the embedding for fast nearest-neighbor search. This is what makes "find anything about our deployment setup" work.

4. 4

   Retrieval on Demand

   When Claude Code asks a question, MemPalace searches for semantically similar memories and injects the most relevant ones into the context window. Under 100ms round-trip.

The result: Claude Code gets the right context at the right time, without you having to repeat yourself. AAAK compression keeps storage efficient at roughly 1MB per 1,000 memories.

Advanced: Connect to ChatGPT & Cursor Too

MemPalace is not limited to Claude Code. The same MCP server works with any compatible client. Once you have MemPalace running, you can connect it to ChatGPT, Cursor, Windsurf, or any other tool that supports MCP. Your memories are shared across all of them — context learned in Claude Code is instantly available in Cursor, and vice versa.

For full configuration details for every supported client, see the complete setup guide.

Frequently Asked Questions