Choose an integration¶
Start with where your application runs and who should own the store.
| Your task | Start here | Where the engine runs |
|---|---|---|
| Give an editor or coding agent persistent memory | MCP and editor setup | A local hms-mcp process, or a shared MCP HTTP server |
| Embed document search in a Rust application | Rust quickstart | Inside your application |
| Add search to a Node application | Node installation | A native Rust module inside Node |
| Access a store from Python or several processes | Python SDK and HTTP server | A server you run; Python is the client |
| Search vectors or experiment with phase hypervectors in Python | Python runtime | The holographic-memory-py extension |
Which Python package?¶
holographic-memory-sdk imports as holographic_sdk. It is an HTTP client for hms-server. The server must be running, and client and server must use the same API key, tenant and embedding dimensions. Installing the SDK does not start a server or download an embedding model.
holographic-memory-py imports as holographic_memory. It exposes native vector search and phase-vector algebra. Choose it when you need those APIs in-process.
Which retrieval path?¶
- Document search: ingest text, then query with words. BM25 works without a model. Add compatible embeddings for semantic retrieval and hybrid search.
- Fact memory: use the memory API to track assertions, replacements and history.
- Editor memory: MCP exposes
remember,recall,searchand other tools to your host. Read the MCP tool reference before enabling write operations.
A similarity score is a ranking signal, not a probability that a statement is true. Keep source text and inspect returned evidence before using it in an answer.
Decide who opens the store¶
An embedded application owns its local store. Multiple clients should connect to one server instead of opening the same writable directory in independent processes. MCP offers stdio for a single client and HTTP for shared access. Its namespace settings determine which memories clients share.
Decide what must stay private¶
The engine's security feature supports encrypted stores. The Python HTTP server has different storage guarantees; see server limitations. Embeddings, sketches and vector masking are not encryption. Read the privacy limits before choosing a deployment.
Continue with installation, then run the complete quickstart.