Neo4j Labs · Agent Memory as a Service

Give your agents a memory that actually remembers.

The Neo4j Agent Memory Service is a hosted, graph-native memory layer for LLM agents - short-term conversations, long-term knowledge, and full reasoning traces - backed by Neo4j Aura with native vector search.

Sign InRead the Docs

Get started in one command

Drop NAMS into your agent CLI with the nams-hooks plugin - memory turns on automatically.

claude plugin marketplace add neo4j-labs/nams-plugins@latest
claude plugin install nams-hooks@nams-plugins

Then set NAMS_API_KEY (and NAMS_BASE_URL=https://memory.neo4jlabs.com) from your dashboard API keys.

See all clients & manual MCP setup →

ClientsAdmin ConsoleDashboard · Memory Browser · Entity ExplorerQuery Console · Observations · API KeysLLM AgentsClaude · Cursor · custom agentsvia MCP (SSE) · memory + workspace toolsApps & SDKsREST clients · backend servicesdirect API-key or Auth0 JWT authNeo4j Agent Memory Service · API Gatewaynams-apiREST API · :8080conversations · entities · tracesnams-mcpMCP / SSE · :9090memory · workspace · skill toolsnams-authAPI keys · Auth0 · JWT · :8081JWKS · scope enforcementnams-tenantsWorkspaces · :8082per-tenant DB provisioningnams-meteringUsage · limits · :8084real-time countersnams-memory-workerAsync: entity extraction · observation & reflectiongeneration · compression · skill distillation → SKILL.mdentity-extractionPython FastAPI · :8085spaCy → GLiNER2 → LLM · embeddingsRedisStreams · counters · cacherate limits · distributed locksNeo4j Aura · Multi-databaseOperational DBworkspaces · users · keysTenant DB · workspace Aisolated graph + vectorsTenant DB · workspace Bisolated graph + vectorsTenant DB · workspace NHNSW cosine indexesBYO Neo4jOptional externalNeo4j instance perworkspaceHTTP Query API v2Short-term · conversationsLong-term · entity graphReasoning · steps & tool callsObservational · reflectionsSemantic search via HNSW · cosine

4 memory types

47 MCP tools

SKILL.md portable skill export

0 infrastructure to run

Everything your agent needs to remember

A complete memory layer - structured, searchable, and production-ready out of the box.

01

Four Memory Types

Short-term conversations, long-term knowledge graph, reasoning traces, and synthesized observations - all in one place.

02

Graph-Native by Design

Typed entities, typed relationships, entity merging, and cross-conversation history. Traversable, queryable, explainable.

03

Semantic Search Built In

HNSW vector indexes on messages, entities, observations and reflections. Hybrid vector + text with automatic fallback.

04

Fully Multi-Tenant

Per-workspace isolated databases on Neo4j Aura, provisioned on demand. Bring-your-own Neo4j also supported.

05

Production Auth & Metering

API keys, Auth0 SSO, scoped JWTs, rate limiting, real-time usage counters - ready for real workloads.

06

Bring Your Own Keys

Per-workspace LLM enrichment with your own provider key (BYOK), or point a workspace at your own external Neo4j - no lock-in.

How it works

Three steps from zero memory to a fully-remembering agent.

STEP 01

Connect

Point your agent at the NAMS MCP endpoint or REST API using your workspace API key.

STEP 02

Remember

Messages, entities, and tool calls are captured, extracted, and embedded automatically.

STEP 03

Recall

Query the graph, search semantically, or ask for the three-tier context for any conversation.

New · Skill Distillation

Turn memory into portable Agent Skills

Distill a scoped slice of a workspace's memory graph into a signed, provenance-grounded SKILL.md package that any agentskills.io runtime can load - every recommendation traces back to the decisions, entities, and messages it came from.

Grounded, not invented

Every claim cites the exact source nodes it came from. The graph is the authority - ungrounded claims are dropped.

Quality gates + human review

Grounding score, coverage, spec lint, and PII redaction must all pass. Nothing publishes without a draft -> in-review -> published sign-off.

Typed procedure graphs

Procedures become a schema-validated step DAG read from the real trace - script-backed steps with typed I/O, not hand-waved prose.

Cryptographic attestation

A dedicated signing key attests each package digest, so consumers can verify offline via JWKS that a skill is unmodified.

Drift detection

A background reconciler re-checks published skills against the live graph and flags stale, missing, or contradicted claims.

Composition & reuse

Shared sub-procedures are lifted into callable skills, so skills compose and reuse instead of duplicating logic.

Seven-stage pipeline

Scope

Extract

Consolidate

Synthesize

Gate

Package

Publish

Explore skill distillation

Stop rebuilding memory from scratch.

NAMS handles extraction, embedding, compression, and retrieval - you keep building the agent.

Sign In to your Workspace