Mnemos
A self-hosted, multi-context memory server that enables AI agents to search and retrieve information from local documents and crawled websites via MCP tools. It runs fully offline using Postgres and Ollama to provide secure, private knowledge management and retrieval-augmented generation.
README
đź§ Mnemos
Self-hosted, Multi-context Memory Server for Developers
Mnemos is an MCP compatible knowledge server that turns your documentation piles into a multi context memory system. It organizes documents into isolated collections, eliminates redundant processing with content hashing, and runs fully offline using Postgres + pgvector and Ollama.
Features
- Multi-context Collections: Isolate your memory by project (e.g.,
react-docs,rust-book,company-internal) with case insensitive search filtering. - Deterministic Re-ingestion: SHA-256 content hashing guarantees idempotent operation—skipping unchanged files and automatically re-chunking on diffs.
- Enhanced Terminal UI: Explore your context with a full screen search interface, result navigation, and detailed chunk inspection modals.
- Recursive Site Crawling: Ingest entire documentation sites with path based filtering (e.g., crawl only
/learnonreact.dev). - Stable Local Embeddings: Optimized for Ollama with persistent connections, automatic runner backoff, and load throttling.
- Chunk Quality Control: Automatic noise filtering (minimum length thresholds + alphanumeric validation) ensures high quality retrieval.
- 100% Private: Fully offline. Your context never leaves your local machine.
Quick Start
Prerequisites
- Docker & Docker Compose
- Python 3.11+
- Ollama (for local embeddings)
1. Install Ollama & Pull Embedding Model
brew install ollama
ollama serve
ollama pull nomic-embed-text
2. Start the Database
cd docker
docker-compose up -d
3. Install Dependencies
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
4. Start the Server
# Option A: Start via CLI (recommended)
python cli/mnemos.py server
# Option B: Run API directly (development)
uvicorn src.main:app --reload
5. Add Documents
python cli/mnemos.py add ./docs/my-document.pdf --collection my-project
# Or crawl a site
python cli/mnemos.py ingest https://react.dev/learn --path-filter /learn --collection react
6. Search
python cli/mnemos.py search "how to use useEffect"
CLI Commands
| Command | Description | Flags |
|---|---|---|
mnemos add <path> |
Add a document or directory | -c <collection>, -r (recursive) |
mnemos ingest <url> |
Ingest a URL or crawl a site | -c <collection>, --path-filter |
mnemos search <query> |
Search for relevant context | -c <collection>, -k <limit> |
mnemos list |
List all documents | -c <collection>, -n <limit> |
mnemos export <file> |
Backup knowledge base to JSON | -c <collection> |
mnemos delete <id> |
Delete a document | -f (force) |
mnemos server |
Start the API server | --host, --port |
API Endpoints
REST API
Mnemos provides a standard REST API for document management and operations.
| Method | Endpoint | Description |
|---|---|---|
POST |
/api/documents |
Upload a document |
GET |
/api/documents |
List all documents |
GET |
/api/collections |
List all unique collections |
GET |
/api/documents/export |
Full JSON backup of chunks |
DELETE |
/api/documents/{id} |
Delete a document |
POST |
/api/search |
Vector similarity search |
POST |
/api/ingest/url |
Ingest a single URL |
POST |
/api/ingest/site |
Crawl a documentation site |
GET |
/api/health |
Health & Stats check |
MCP Endpoints
Mnemos exposes its retrieval capabilities via the Model Context Protocol (MCP), allowing AI agents to query it as an external context provider. Mnemos is designed to be stateless from the MCP client’s perspective; all persistence lives server-side.
| Method | Endpoint | Description |
|---|---|---|
GET |
/mcp/tools |
List available MCP tools |
POST |
/mcp/call |
Execute an MCP tool |
MCP Integration
Claude Desktop
Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"mnemos": {
"command": "curl",
"args": ["-X", "POST", "http://localhost:8000/mcp/call", "-H", "Content-Type: application/json", "-d"]
}
}
}
Available MCP Tools
- search_context: Search the knowledge base for relevant context
- list_documents: List all documents in the knowledge base
- get_document_info: Get detailed information about a document
Configuration
Environment variables (.env):
| Variable | Default | Description |
|---|---|---|
DATABASE_URL |
postgresql+asyncpg://... |
Postgres connection string |
EMBEDDING_PROVIDER |
ollama |
ollama (local-first default) or openai |
EMBEDDING_MODEL |
nomic-embed-text |
Ollama embedding model |
OLLAMA_BASE_URL |
http://127.0.0.1:11434 |
Ollama API URL |
CHUNK_SIZE |
300 |
Target characters per chunk |
CHUNK_OVERLAP |
40 |
Overlap between chunks |
Architecture
graph TD
User([User CLI / App]) --> API[FastAPI Server]
API --> DB[(PostgreSQL + pgvector)]
API --> Ollama[Ollama Local Embeddings]
subgraph Ingestion Pipeline
API --> Parser[Document Parser]
Parser --> Chunker[Text Chunker]
Chunker --> HashCheck[SHA-256 Content Hash]
HashCheck --> Embedding[Vector Generation]
end
subgraph Retrieval
API --> Search[Vector Search]
Search --> Context[Context Assembler]
end
Design Principles
- Local-first by default: All heavy lifting (vectors/search) happens on your hardware.
- Deterministic ingestion: SHA-256 hashing ensures idempotency and safe re-runs.
- Explicit context isolation: Multi-collection support prevents cross-project context pollution.
- Inspectable retrieval: Similarity scores and chunk metadata are exposed to build trust.
- Zero vendor lock-in: Standards-based tech stack (Postgres, MCP, REST).
Supported Embedding Models
| Model | Dimensions | Notes |
|---|---|---|
nomic-embed-text |
768 | Default, good balance |
mxbai-embed-large |
1024 | Higher quality |
all-minilm |
384 | Faster, smaller |
Security Posture
- Local-Only: By default, Mnemos binds to
0.0.0.0but does not include authentication. It is intended for local use or behind a secure tunnel. - No External Calls: All vector generation and retrieval happen locally. No telemetry or document data is sent to external servers.
- SQLi Prevention: Uses SQLAlchemy ORM and parameterized queries for all database interactions.
Non-Goals
- Cloud Hosting: Mnemos is not designed to be a multi-tenant cloud SaaS.
- Advanced LLM Orchestration: It focuses on context provision, not on being a full RAG agent.
- Browser Automation: Ingestion is via CLI or URL crawler, not a GUI automation tool.
Development
black src/ cli/
pytest tests/
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.