AutoGineer
Generated banner for Memorizer
Generated logo for Memorizer

Memorizer

Memory serverOfficial Petabridge project

Created by Dario Griffo

Published in petabridge/memorizer. Publisher page

221 GitHub stars

What it is

Vector-search powered agent memory MCP server
— the description by Dario Griffo

AutoGineer overview, from the brain’s own files

Memorizer is a dedicated memory service designed to give AI agents persistent, long-term storage capabilities. It functions as an MCP server, allowing agents to store, retrieve, and semantically search through structured memories. Built on .NET and PostgreSQL with the pgvector extension, it supports advanced features like versioning, audit trails, relationship mapping for knowledge graphs, and workspace-based organization. It includes a web-based UI for manual management and provides a robust framework for agents to proactively manage their own knowledge base, ensuring that insights and context are preserved across sessions.

How it works

Memory model: Hybrid

AutoGineer overview, from the brain’s own files

Memorizer runs as a service that connects to an AI agent via the Model Context Protocol. It stores data in a PostgreSQL database, utilizing vector embeddings for semantic similarity searches. The system organizes information into workspaces and projects, supports metadata tagging, and tracks version history for all memory entries. Agents interact with the service through a set of defined tools for CRUD operations, metadata updates, and relationship creation, enabling the construction of an interconnected knowledge graph.

Folders
FolderRoleFiles
.githubagent-config4
docsdocs37
scriptsscripts3
srcsource-code195
.configfolder1

The agent writes

Requirements

Databases: PostgreSQL

Works with: Claude Code, OpenCode, Devin (ex-Windsurf)

Set it up

  1. Not scanned by AutoGineerThe author’s README also describes installing it with Docker Compose. Read its install section: AutoGineer did not scan that path. https://github.com/petabridge/memorizer/blob/d3e90fa8d6b50254d65a3ffa3acc35e0204c5c90/README.md#-quick-start-public-image

  2. Created by Dario Griffo. Published in petabridge/memorizer. Licence: MIT License.

Usage

AutoGineer overview, from the brain’s own files

  • Connects to AI agents via the Model Context Protocol (MCP) to provide persistent memory.
  • Enables semantic search to retrieve relevant context based on task descriptions.
  • Supports proactive memory management where agents store, update, and relate information autonomously.
  • Provides a web interface for human users to manage, audit, and organize memory entries.
  • Facilitates knowledge graph construction by creating explicit relationships between memory nodes.
“Search for previous instructions on how to configure the build pipeline.”
The agent queries the memory server using semantic similarity to find relevant 'how-to' notes and returns the specific configuration steps.
“Store this new insight about the project architecture as a reference.”
The agent saves the markdown content to the memory server with appropriate tags and metadata for future retrieval.
“Link this new task to the related project requirements memory.”
The agent creates a relationship edge between the new memory ID and the existing requirements memory ID.

Documentation by Dario Griffo

Memorizer

Image: Docker Pulls Image: GitHub License Image: GitHub Actions Workflow Status Image: GitHub Release

Memorizer is a .NET-based service that allows AI agents to store, retrieve, and search through memories using vector embeddings. It leverages PostgreSQL with the pgvector extension to provide efficient similarity search capabilities.

Image: Memorizer - Memory List

Key features:

  • Workspaces & Projects - Organize memories into hierarchical workspaces and projects with status tracking

  • Store structured memories with vector embeddings

  • Retrieve memories by ID

  • Semantic search through memories using vector similarity

  • Filter search results using tags

  • Edit memory content with automatic versioning and change tracking

  • Update memory metadata (title, type, tags, confidence) independently

  • Revert memories to previous versions with full audit trail

  • Create relationships between memories to form knowledge graphs

  • Web UI for manually adding, editing, deleting, viewing memories, and managing versions

  • Provider Settings - Configure embedding and LLM providers through the UI (Ollama, OpenAI, and compatible APIs)

  • MCP (Model Context Protocol) integration for easy use with AI agents

  • Light & Dark theme support

Image: Memorizer - Workspace Detail

Technologies

  • .NET 10.0

  • PostgreSQL with pgvector extension

  • Model Context Protocol (MCP)

  • ASP.NET Core

  • Akka.NET for background jobs, such as re-embedding memories if you change algorithms

  • Npgsql for PostgreSQL connectivity

📖 Documentation

Installation with Docker

🐳 Quick Start (Public Image)

The easiest way to get started is using the pre-built Docker image and our docker-compose.yml file:

docker-compose up -d

This will:

View the Memorizer Web UI on http://localhost:5000.

🚀 Local Development Builds

If you want to build and run from source:

Prerequisites

  • Docker and Docker Compose

  • .NET 10.0 SDK

1. Build and Publish Local Container
# From solution root directory
# Build and publish the .NET container
dotnet publish -c Release /t:PublishContainer

This creates a container image named memorizer:latest.

2. Start Infrastructure and Application
docker-compose -f docker-compose.local.yml up -d

This starts the same services but uses your locally built image.

Upgrading to Memorizer 2.0

NOTE
The upgrade from Memorizer 1.x to 2.0 is designed to work automatically. All schema migrations run on startup and have been thoroughly tested. Your existing memories will be preserved and continue to work as expected.

CAUTION
We recommend backing up your database before upgrading, as a standard best practice. Memorizer 2.0 runs automatic schema migrations on startup that alter tables and add new columns. While we've tested these migrations extensively, having a backup ensures you can recover in the unlikely event that something goes wrong.

Backup Instructions

NOTE
The container names below (e.g., memorizer-postgres) are based on the default docker-compose.yml. If you've customized your Docker Compose file, adjust the container name accordingly.

Before pulling the new Memorizer 2.0 image, create a PostgreSQL dump of your existing database:

# Create a full database backup
docker exec postgmem-postgres pg_dump -U postgres postgmem > memorizer-backup-$(date +%Y%m%d).sql

# Or use pg_dump with compression
docker exec postgmem-postgres pg_dump -U postgres -Fc postgmem > memorizer-backup-$(date +%Y%m%d).dump

To restore from a backup if needed:

# Restore from SQL dump
docker exec -i postgmem-postgres psql -U postgres postgmem < memorizer-backup-20250207.sql

# Or restore from compressed dump
docker exec -i postgmem-postgres pg_restore -U postgres -d postgmem memorizer-backup-20250207.dump
Updating
# make sure you have created a backup before starting here
git pull
docker-compose down
docker-compose pull memorizer
docker-compose up -d
Breaking Changes in 2.0

  • MCP endpoint moved from / to /mcp. Update your MCP client configurations from http://localhost:5000 to http://localhost:5000/mcp.

  • Web UI moved from /ui to /. The Web UI is now at the root path.

  • New database tables: Workspaces, projects, provider settings, data migrations, and archetype tracking are added automatically on first startup.

  • Existing memories are preserved: All your V1 memories will continue to work. They will appear as "Unfiled" until you organize them into workspaces and projects.

After Upgrading

TIP
Once the upgrade is complete, ask your AI agent to suggest workspaces and projects to organize your existing memories into. Memorizer 2.0's workspace and project system helps you keep memories structured and easy to find - and your agent can analyze your existing memories to recommend a good organizational scheme.

🔌 MCP Configuration Example

To use Memorizer with any MCP-compatible client, add the following to your configuration (e.g., mcp.json):

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

This uses the modern Streamable HTTP transport (MCP spec 2025-03-26+). Note that the MCP endpoint is at /mcp (changed from / in 2.0).

🖥️ Web UI

Memorizer includes a web-based user interface for managing memories through your browser.

Access the Web UI

Once the application is running (via docker-compose up -d), you can access the Web UI at:

http://localhost:5000/

Features

  • Memory Management: Create, view, edit, and delete memories with full CRUD operations

  • Version Control: View version history, compare versions with visual diffs, and revert to previous versions

  • Search & Filter: Search memories using semantic similarity and filter by tags

  • Statistics Dashboard: View memory counts, tag distributions, and system statistics

  • MCP Configuration: Get the MCP configuration JSON for connecting clients at /mcp-config

The Web UI provides a user-friendly interface for all Memorizer functionality, making it easy to manage your AI agent's memory without needing to use the MCP tools directly.

🧠 Example System Prompt for LLMs

IMPORTANT
⚡ Pro Tip: Add this system prompt to your AGENT.md, Cursor Rules files, or any AI agent configuration! This will dramatically improve how often and effectively your LLM uses the Memorizer service for persistent memory management.

You have access to a long-term memory system via the Model Context Protocol (MCP) at the endpoint memorizer. Use the following tools:

Storage & Retrieval:

  • store: Store a new memory. Parameters: type, text (markdown), source, title, tags, confidence, relatedTo (optional, memory ID), relationshipType (optional).

  • searchMemories: Search for similar memories using semantic similarity. Parameters: query, limit, minSimilarity, filterTags.

  • get: Retrieve a memory by ID. Parameters: id, includeVersionHistory, versionNumber.

  • getMany: Retrieve multiple memories by their IDs. Parameter: ids (list of IDs).

  • delete: Delete a memory by ID. Parameter: id.

Editing & Updates:

  • edit: Edit memory content using find-and-replace (ideal for checking off to-do items, updating sections). Parameters: id, old_text, new_text, replace_all.

  • updateMetadata: Update memory metadata (title, type, tags, confidence) without changing content. Parameters: id, title, type, tags, confidence.

Relationships & Versioning:

  • createRelationship: Create a relationship between two memories. Parameters: fromId, toId, type (e.g., 'example-of', 'explains', 'related-to').

  • revertToVersion: Revert a memory to a previous version. Parameters: id, versionNumber, changedBy.

All edits and updates are automatically versioned, allowing you to track changes and revert if needed. Use these tools to remember, recall, edit, relate, and manage information as needed to assist the user.

License

MIT

💖 Attribution

Made with ❤️ by Petabridge

Originally forked from Dario Griffo's postg-mem server

docs/embedding-migration.md

Embedding Dimension Migration

This document explains how Memorizer handles changes to embedding dimensions when you switch to a different embedding model.

Overview

Memorizer uses pgvector for storing and searching vector embeddings. pgvector uses fixed-dimension VECTOR(n) columns, which means:

  • All embeddings in a column must have the same number of dimensions

  • Changing dimensions requires schema migration (ALTER TABLE)

  • Existing embeddings become invalid when dimensions change

Common Embedding Models and Dimensions

ModelDimensionsNotes
all-minilm384Fast, lightweight, good for general use
nomic-embed-text768Higher quality, more compute required
mxbai-embed-large1024High quality, large model
bge-base-en-v1.5768BGE family, good quality
bge-large-en-v1.51024BGE family, high quality

How Auto-Detection Works

Memorizer automatically detects embedding dimensions:

  1. On Startup: Probes the configured embedding model by generating a test embedding

  2. Stores Config: Records the detected model name and dimensions in embedding_config table

  3. Validates: Compares detected dimensions against stored config and database schema

  4. Warns: If mismatch detected, shows warning banner in UI

Sources of Truth

  1. Detected Dimensions: Live probe of the embedding API (most accurate)

  2. Stored Config: Previous model/dimension recorded in embedding_config table

  3. Schema Dimensions: Actual VECTOR(n) column size in database

When Migration is Needed

Migration is required when:

  • You change Embeddings:Model in configuration to a model with different output dimensions

  • The detected dimensions don't match the database schema

  • Example: Switching from all-minilm (384d) to nomic-embed-text (768d)

Migration Process

The dimension migration tool performs these steps:

  1. Acquire Lock: Uses PostgreSQL advisory lock to prevent concurrent migrations

  2. Drop Indexes: Removes vector similarity indexes

  3. ALTER TABLE: Changes VECTOR(n) column to new dimension size

    • Sets all existing embeddings to NULL (search temporarily unavailable)

  4. Recreate Indexes: Rebuilds vector indexes for new dimensions

  5. Regenerate Embeddings: Re-embeds all memories with new model

  6. Update Config: Stores new model/dimension in embedding_config

  7. Release Lock: Allows normal operations to resume

What Happens to Old Embeddings?

During migration, old embeddings are set to NULL because:

  • pgvector doesn't support zero-padding or truncating vectors

  • Old dimension vectors are incompatible with new dimensions for similarity search

  • Setting to NULL ensures searches don't return invalid results

Memories with NULL embeddings won't appear in vector similarity searches until regeneration completes.

Using the Migration Tool

Via UI

  1. Navigate to Tools > Dimension Migration

  2. Review the current dimension status

  3. If migration is needed, click Start Dimension Migration

  4. Monitor progress via the real-time progress bar

Via API
# Check dimension status
curl http://localhost:5000/ui/tools/dimension-status

# Start migration
curl -X POST http://localhost:5000/ui/tools/start-dimension-migration

# Check migration status
curl http://localhost:5000/ui/tools/dimension-migration-status

Configuration

Embedding configuration is in appsettings.json:

{
  "Embeddings": {
    "ApiUrl": "http://localhost:11434/",
    "Model": "all-minilm",
    "Timeout": "00:00:10"
  }
}

Note: The Dimensions setting has been removed. Dimensions are now auto-detected from the model.

Resuming Failed Migrations

If a migration is interrupted, Memorizer tracks the migration state in the embedding_dimension_migrations table. You can resume from the UI or via API:

curl -X POST http://localhost:5000/ui/tools/resume-dimension-migration \
  -H "Content-Type: application/json" \
  -d '{"migrationId": "your-migration-id"}'

Monitoring

  • Startup Logs: Check for "Validating embedding dimensions" messages

  • Warning Banner: Appears in UI when mismatch detected

  • Migration Table: Query embedding_dimension_migrations for history

Troubleshooting

"Embedding API unavailable"

  • Ensure Ollama (or your embedding service) is running

  • Check Embeddings:ApiUrl in configuration

  • Verify the model is pulled: ollama pull your-model-name

"Could not acquire migration lock"

  • Another migration may be in progress

  • Check for stuck locks: SELECT * FROM pg_locks WHERE locktype = 'advisory'

  • If needed, terminate the blocking connection

Search not returning results during migration

  • This is expected - embeddings are being regenerated

  • Wait for migration to complete

  • Check progress in the Tools > Dimension Migration page

Database Schema

The migration system uses these tables:

-- Tracks current embedding configuration
CREATE TABLE embedding_config (
    id SERIAL PRIMARY KEY,
    model_name TEXT NOT NULL,
    dimensions INT NOT NULL,
    detected_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    is_active BOOLEAN NOT NULL DEFAULT true
);

-- Tracks migration history
CREATE TABLE embedding_dimension_migrations (
    migration_id UUID PRIMARY KEY,
    started_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    completed_at TIMESTAMPTZ,
    old_model TEXT NOT NULL,
    old_dimensions INT NOT NULL,
    new_model TEXT NOT NULL,
    new_dimensions INT NOT NULL,
    status TEXT NOT NULL,  -- 'running', 'schema_changed', 'regenerating', 'completed', 'failed'
    total_memories INT NOT NULL DEFAULT 0,
    memories_processed INT NOT NULL DEFAULT 0,
    memories_successful INT NOT NULL DEFAULT 0,
    memories_failed INT NOT NULL DEFAULT 0,
    error_message TEXT,
    requested_by TEXT
);
260 files · 244 not served here

Advantages

  • Provides persistent, long-term memory that survives agent session resets. evidence: README.md
  • Supports semantic search through vector embeddings, allowing for context-aware retrieval. evidence: README.md
  • Includes full versioning and audit trails for all memory modifications. evidence: README.md
  • Allows for the creation of knowledge graphs by linking related memories. evidence: README.md

Limitations

  • Requires external infrastructure including PostgreSQL with pgvector and an embedding provider. evidence: README.md
  • Requires manual configuration of MCP clients to point to the correct endpoint. evidence: README.md
  • Upgrades may require database backups and potential schema migrations. evidence: README.md

Security

Risk level: No risk found

This brain runs

This brain runs nothing: AutoGineer found no hooks, scripts, plugin code or MCP servers in its files.

Attribution and licence

Created by
Dario Griffo
Publisher
Petabridge (Official Petabridge project)
Source
petabridge/memorizer · commit d3e90fa
Licence
MIT License (files served here)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by Dario Griffo or Petabridge.

Install from source

244 of its files are not served here, so there is no ZIP. petabridge/memorizer at d3e90fa