mcp-memory-ydb
An official long-term memory MCP server for agents on YDB Serverless: mem0 extracts facts, YDB stores the vectors
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Stores extracted facts about the user in a YDB database readable by anyone with access to that database
- Sends text to a third-party LLM provider for fact extraction and embeddings
Install
Manual install
uvx mcp-memory-ydb setupThe README's setup wizard: creates .mcp-memory-ydb.env and prints ready connection commands.
This is third-party code. Review the repository files before installing.
What it does
The server is built on ydb-mcp, Yandex's own official MCP server for YDB, with the generic SQL tools switched off and four memory tools in their place: memory_search does a semantic lookup over facts, memory_save hands text to mem0, which uses an LLM to extract facts and append them (facts are not auto-overwritten, only accumulated), and memory_delete / memory_update edit a specific fact by its id. All memory lives in YDB Serverless as a vector store via langchain-ydb, and fact extraction plus embeddings work with any OpenAI-compatible provider, OpenAI, YandexGPT or local Ollama.
Who it is for. For agent developers who want cross-session long-term memory on their own YDB infrastructure rather than a third-party cloud.
Good fit when
- You need agent memory that survives a session restart, built on YDB
- You already have, or are willing to set up, YDB Serverless with a service account holding the ydb.editor role
- You want facts stored in Russian without being translated to English
Not a fit when
- You have no access to YDB Serverless or an OpenAI-compatible LLM for fact extraction
- You need multi-tenant memory with access control: a namespace is a query filter, not a security boundary
Example request
Remember that my favorite language is Rust and I work in the Moscow timezoneLimitations
mem0 2.x is append-only: it does not auto-delete or merge facts, stale facts need manual cleanup via memory_delete/update. A namespace is not a security boundary: anything with access to the YDB database can read every namespace. By default extracted facts are stored in English; for Russian you need to set MEMORY_FACT_INSTRUCTIONS in Russian.
How to disable. Run claude mcp remove memory-ydb, or remove the memory-ydb entry from your MCP client config.
MCP
- Transport
- stdio
- Authentication
- API key
| Environment variables | |
|---|---|
| YDB_ENDPOINT required | The YDB Serverless endpoint, e.g. grpcs://ydb.serverless.yandexcloud.net:2135. |
| YDB_DATABASE required | The database path, /ru-central1/<folder>/<db>. |
| YDB_SA_KEY_FILE secret | Path to the service account's JSON key with the ydb.editor role. |
| LLM_API_KEY required, secret | API key for an OpenAI-compatible provider (OpenAI, Yandex Cloud, Ollama) used for fact extraction and embeddings. |
Security check
- Stores extracted facts about the user in a YDB database readable by anyone with access to that database
- Sends text to a third-party LLM provider for fact extraction and embeddings
README in short
The README explains the memory model in detail (mem0 extracts facts, langchain-ydb stores the vectors, YDB Serverless is the database), gives a uvx mcp-memory-ydb setup wizard, connection instructions for Claude Code, Cursor and VS Code, a section explaining that a namespace is a partition label rather than a security boundary, and a troubleshooting section for common YDB connection errors.
FAQ
Do I need to call memory_search and memory_save by hand?
No, the server ships built-in MCP instructions, and clients like Claude Code call them automatically before and after answering.
Can I use YandexGPT instead of OpenAI for embeddings?
Yes, any OpenAI-compatible provider works, including Yandex Cloud.
Related
166 skills for scientific work: bioinformatics, cheminformatics, clinical data, geospatial analysis and 100+ databases
Google's open-source MCP server for databases: ready tools for Postgres, MySQL, BigQuery, Spanner and more, plus custom tools in tools.yaml
Official Hugging Face skills: Hub operations via the hf CLI, datasets, model training, Spaces, evals and deployment
A visualization language for agents and an MCP server: neat charts from a simple semantic spec