Vector Memory MCP Server

Semantic memory storage for AI assistants. Store decisions, patterns, and context that persists across sessions.

A local-first MCP server that provides vector-based memory storage. Uses local embeddings and SQLite with sqlite-vec for fast, private semantic search — all in a single file.

Features

• Local & Private - All embeddings generated locally (all-MiniLM-L6-v2, 384-dim), data stored in a single SQLite file

• Semantic Search - Hybrid vector + full-text search with intent-based ranking

• Batch Operations - Store, update, delete, and retrieve multiple memories at once

• Session Waypoints - Save and restore project context between sessions

• Conversation History - Index and search Claude Code session transcripts

• MCP Native - Standard protocol, works with any MCP-compatible client

Quick Start

Prerequisites

• Bun 1.0+

• An MCP-compatible client (Claude Code, Claude Desktop, etc.)

Install

bun install -g @aeriondyseti/vector-memory-mcp

First install downloads ML models (~90MB). This may take a minute.

Configure

Add to your MCP client config (e.g., ~/.claude/settings.json):

{
  "mcpServers": {
    "vector-memory": {
      "type": "stdio",
      "command": "bunx",
      "args": ["--bun", "@aeriondyseti/vector-memory-mcp"]
    }
  }
}

Use

Restart your MCP client. You now have access to:

Usage

Store a memory:

You: "Remember that we use Drizzle ORM for database access"
Assistant: [calls store_memories]

Search memories:

You: "What did we decide about the database?"
Assistant: [calls search_memories with relevant query]

Session waypoints:

You: "Save context for next session"
Assistant: [calls set_waypoint with summary, completed items, next steps]

Conversation history (requires --enable-history):

You: "What did we discuss about the API design last week?"
Assistant: [calls search_memories with history_only: true, history_before/after filters]

Configuration

CLI flags:

Release Channels

The stable release is what you get by default:

bun install -g @aeriondyseti/vector-memory-mcp

Pre-release channels are available for testing upcoming changes. These are unstable and may break without notice — use at your own risk.

# Install the dev channel
bun install -g @aeriondyseti/vector-memory-mcp@dev

# Pin to a specific pre-release version
bun install -g @aeriondyseti/vector-memory-mcp@2.1.0-dev.1

# Go back to stable
bun install -g @aeriondyseti/vector-memory-mcp@latest

Warning: Pre-release versions may include breaking changes, incomplete features, or data migration requirements that haven't been finalized. Do not use them in production workflows you depend on.

Migrating from 1.x (LanceDB)

Version 2.0 replaced LanceDB with SQLite (sqlite-vec) for storage. If you have existing data from 1.x, the server will detect it automatically and prompt you to migrate:

vector-memory-mcp migrate

This reads your LanceDB directory, writes a new SQLite file, and prints instructions to swap them. Your original data is preserved until you manually remove it.

What changed:

• Storage: LanceDB directory (~845 files) → single .db file

• Dependencies: 223MB (@lancedb/lancedb + apache-arrow) → 24KB (sqlite-vec)

• Runtime: Node.js support dropped, Bun required (for bun:sqlite)

Development

git clone https://github.com/AerionDyseti/vector-memory-mcp.git
cd vector-memory-mcp
bun install

bun run test      # Run all tests
bun run dev       # Watch mode
bun run typecheck # Type checking

See CHANGELOG.md for release history and ROADMAP.md for planned features.

Contributing

Contributions welcome! See issues for areas we'd love help with.

License

MIT - see LICENSE

Built with MCP SDK, sqlite-vec, and Transformers.js