Deploy a memory application

HTTPS REST, https://memory.neo4jlabs.com/v1 by default

NAMS (no database to operate)

A workspace API key (MEMORY_API_KEY, prefixed nams_), plus MEMORY_WORKSPACE_ID (sent as X-Workspace-Id) on deployments that scope by header rather than by key; a workspace-bound key rejects a different workspace id

No Neo4j instance or extraction pipeline to run yourself. Some Bolt-only operations (preferences, facts, client-created relationships, geospatial search, custom Cypher writes) are unsupported; see backend capabilities. See Connect a Python application to NAMS.

Bolt (bolt://, neo4j+s://) from your own process — no MCP server

Bolt, against a Neo4j instance you operate (self-managed or AuraDB)

NAM_NEO4JURI / NAM_NEO4JUSERNAME / NAM_NEO4J__PASSWORD, which MemorySettings() reads from the environment, plus embedding/LLM provider credentials for the selected providers (MemorySettings implicitly selects OpenAI embeddings). Plain NEO4J_URI / NEO4J_USERNAME / NEO4J_PASSWORD are not read by MemorySettings; an application that uses them must pass the values itself, for example MemorySettings(neo4j=…​). See environment variables.

Full Bolt operation set, including preferences, facts, client-created relationships, geospatial search and custom Cypher; you operate the database and the extraction pipeline. See configuration reference.

stdio (the CLI default)

Bolt, against a Neo4j instance you operate, or NAMS (--backend nams, the default when MEMORY_API_KEY is set)

For Bolt, NEO4J_URI / NEO4J_USER / NEO4J_PASSWORD (the mcp serve options read these names); for NAMS, MEMORY_API_KEY (and MEMORY_ENDPOINT to override the default endpoint). Provider credentials (for example an OpenAI key) only if extraction or embeddings are enabled

For local MCP hosts — Claude Desktop, Claude Code, Cursor. Not a network-reachable deployment: the host process owns stdio.

http (streamable-http is an explicit spelling of the same transport; sse is accepted but deprecated and served as http)

Bolt, against a Neo4j instance you operate (for example AuraDB), or NAMS (--backend nams, the default when MEMORY_API_KEY is set)

For Bolt, NEO4J_URI / NEO4J_USER / NEO4J_PASSWORD; for NAMS, MEMORY_API_KEY. Plus embedding/LLM provider credentials for whichever provider you select

The documented example (Cloud Run MCP deployment) runs this transport on Cloud Run with Vertex AI embeddings (gemini-embedding-001, 768 dimensions), Secret Manager for NEO4J_URI/NEO4J_USER/NEO4J_PASSWORD, Cloud Run IAM for endpoint authentication, and automatic extraction/preference detection disabled by default. It has no built-in application-level authentication — the platform (Cloud Run IAM, or your own reverse proxy) must supply that.

HTTPS REST, via fetch only (zero runtime dependencies)

NAMS

MEMORY_API_KEY, passed explicitly through the runtime’s request-scoped bindings (env.MEMORY_API_KEY on Workers) rather than read from process.env

Bolt is not reachable from these runtimes; NAMS/REST is the only supported backend here. See Deploy to edge runtimes for the API-key-resolution gotcha and what does and doesn’t work on edge (for example, the TCK bridge transport is not tested there).