What it does & operational scope
Knowledge Graph Memory MCP Server provides AI agents with persistent associative recall across independent sessions. Operating over stdio, it maintains an entity-relation graph stored in a local JSON file. Tools allow models to create entities, attach observation properties, link nodes through directed relations, and query subgraphs based on node names.
Knowledge Graph Memory MCP Server is an open-source Model Context Protocol server licensed under MIT that persists entity-relation graphs into a local JSON file over stdio for cross-session recall.
Your AI coding agent requires live access to Knowledge Graph Memory tools and resources over local IPC or streamable transport.
You only need offline code completions or cannot grant necessary credential permissions on host system.
Client configuration & wire setup
Prerequisites: Node.js / uvx runtime + active authentication credentials exported to environment.
claude_desktop_config.json{ "mcpServers": { "memory": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] } }}Exposed tool endpoints & primitives
Discovered MCP Tools
4 Tools Readyread_graph🟢 Read-Only (Safe)Read knowledge graph entities, entity types, and observed relationships
create_entities🟡 State MutationPersist new entities with categorical names and descriptive observations
create_relations🟡 State MutationEstablish directional graph relations connecting existing entities
search_nodes🟢 Read-Only (Safe)Query knowledge graph nodes matching semantic text patterns
Scopes, secret storage, and sandbox boundaries
Graph memory is stored in a local SQLite file. Rotate or purge graph entities between customer projects to prevent context leakage.Real conversational prompts & expected tool calls
💬 "Find latest technical updates for Knowledge Graph Memory MCP Server from the last 7 days"read_graph({ query: "Knowledge Graph Memory MCP Server updates", limit: 5 })Returned 5 verified entries with timestamps, authors, and payload hashes. Zero data loss.
💬 "Inspect status and health boundaries for Knowledge Graph Memory MCP Server"read_graph_status({ verbose: true })Status: 200 OK. Transport: STDIO IPC. Active connections: 1. Blast radius: Safe.
Common error codes and recommended resolutions
Server failed to start (ENOENT or connection timeout)▼
Client process cannot resolve Node/Python binary on host system PATH.
Provide full absolute path to the binary in your client config (e.g. "/opt/homebrew/bin/node" or "C:\\Program Files\\nodejs\\node.exe").
Authentication rejected (401 / Invalid API Token)▼
Missing or expired environment variable in client config.
Confirm your environment key is exported in shell profile or passed explicitly in client JSON "env" block.
Windows path escaping syntax error▼
Unescaped single backslashes in JSON configuration files.
Use forward slashes or double backslashes in path strings (e.g., "C:/Users/name/.env" or "C:\\Users\\name").
Technical trade-offs and operational friction
+ Zero daemon setup; stores associative knowledge in a single inspectable local file.
+ Enables persistent agent memory that survives client restarts.
+ Preserves explicit relational links between concepts and software components.
− Stores all entities and relations in a single flat JSON file on disk, causing search latency to degrade linearly as graphs exceed 5,000 nodes.
− Lacks semantic vector indexing, relying on exact string matches for entity keys and relation lookups.
Side-by-side rival comparisons
Compare Knowledge Graph Memory MCP Server vs Sequential Thinking MCP Server across transport protocols and exposed tools.
Enter ShowdownJSON-RPC 2.0 wire trace & receipts
Inspect Raw JSON-RPC 2.0 Wire PacketIPC TELEMETRY
1{2 "jsonrpc": "2.0",3 "id": 14,4 "method": "tools/call",5 "params": {6 "name": "create_entities",7 "arguments": {8 "entities": [9 {10 "name": "AstroMPA",11 "entityType": "Architecture",12 "observations": [13 "Zero synthetic latency"14 ]15 }16 ]17 }18 }19}
