McpKnowledgeGraph
title: mcp-knowledge-graph radar_quadrant: Tools radar_ring: Assess radar_position: inner created: 2026-05-26 last_updated: 2026-05-26 tags: [mcp, memory, knowledge-graph, agents, local-first] source_url: https://github.com/shaneholloman/mcp-knowledge-graph
A local-first MCP server that gives AI agents persistent memory via a local knowledge graph. Facts are stored as typed nodes and edges on the local machine; the agent can query and traverse relationships between entities across sessions. A community fork of the official Anthropic memory MCP server, optimised for local development with no cloud dependency.
Most agent memory implementations store key-value facts or free-text notes. A knowledge graph stores entities and the relationships between them — "Project X uses Framework Y" and "Framework Y was created by Organisation Z" are traversable links, not isolated records. An agent querying "what does Project X depend on?" can follow edges rather than keyword-matching across flat notes.
The fork's primary differentiation is local storage. The official memory-mcp server targets managed deployment; this fork persists the graph to local files, making it suitable for personal workstations, air-gapped environments, and development setups where cloud memory storage is not appropriate.
Complements Context7 in the MCP server ecosystem — context7 injects current docs at query time; mcp-knowledge-graph persists cross-session agent memory. Relevant to BuildingEffectiveAgents and LLMSystemPatterns as the memory layer for agentic workflows. The knowledge-graph structure complements Txtai (all-in-one AI framework with graph search).
Placed in Tools / Assess / inner. Local-first persistent agent memory with graph structure is meaningfully distinct from flat file memory and cloud-dependent solutions. Inner position reflects applicability to any team building agents that need to retain and traverse context across sessions without cloud infrastructure. Trial gate: agent using the knowledge graph across at least three sessions with relationship traversal confirmed producing relevant recall.
