Memorizer

Memorizer is a .NET-based memory service that AI agents can use via the Model Context Protocol (MCP) to store, retrieve, and semantically search memories using vector embeddings. It leverages PostgreSQL with pgvector to provide fast similarity search and supports versioning, tagging, relationships, and a Web UI for manual management. You can connect to Memorizer over HTTP with MCP or run a local stdio instance for tighter integration.

How to use

You connect to Memorizer from an MCP client using the Memorizer HTTP endpoint. The service stores memories as structured records with vector embeddings and exposes operations to store new memories, search by similarity, retrieve by ID, edit content, update metadata, and manage memory relationships. Use the MCP facade to build long-term memory flows for your AI agent: store important facts, search for relevant memories, and link related memories to form a knowledge graph. All edits are versioned, so you can audit changes and revert when needed.

How to install

Prerequisites: ensure you have Docker and Docker Compose installed, and you may also want the .NET 9.0 SDK if you plan to build locally.

# Quick start using the public image with Docker Compose
docker-compose up -d

# The stack starts PostgreSQL with pgvector, the web UI, and the Memorizer API
# Access the Web UI at http://localhost:5000/ui

If you prefer to run from source for local development, follow the local development flow. Build and publish a local container image, then start the infrastructure with a local compose file.

# From the solution root, publish a local container image
dotnet publish -c Release /t:PublishContainer

# Start infrastructure using the local compose file
docker-compose -f docker-compose.local.yml up -d

Configuration and MCP connection

Memorizer exposes an MCP HTTP endpoint you can connect to from clients. Use the following MCP configuration snippet to point your client at the Memorizer service.

{
  "memorizer": {
    "url": "http://localhost:5000"
  }
}

Available tools

store

Store a new memory with type, text, source, title, tags, and optional metadata like confidence and relatedTo.

searchMemories

Search for memories using semantic similarity with parameters like query, limit, minSimilarity, and filterTags.

get

Retrieve a memory by ID, with optional version history.

getMany

Retrieve multiple memories by their IDs.

delete

Delete a memory by ID.

edit

Edit memory content with find-and-replace support and optional diffing.

updateMetadata

Update memory metadata such as title, type, tags, and confidence without changing content.

createRelationship

Create relationships between memories to build knowledge graphs.

revertToVersion

Revert a memory to a previous version with full audit trail.