MemoryMcp
title: memory-mcp radar_quadrant: Tools radar_ring: Assess radar_position: inner created: 2026-05-26 last_updated: 2026-05-26 tags: [mcp, memory, agents, anthropic, persistence, key-value] source_url: https://github.com/modelcontextprotocol/servers/tree/main/src/memory
The official Anthropic MCP server for persistent agent memory, part of the modelcontextprotocol/servers reference implementation. Stores facts as a flat key-value structure in a local JSON file, enabling Claude and other MCP-compatible agents to recall information across sessions. The reference baseline for agent memory in the MCP ecosystem.
Between sessions, the agent can save named facts ("user_preferred_language: Python", "project_deadline: 2026-06-01") and retrieve them on demand. Storage is a plain JSON file on the local machine. The MCP interface exposes create_entities, create_relations, add_observations, delete_entities, search_nodes, and read_graph tools — making the memory queryable as a basic knowledge graph, not purely key-value.
McpKnowledgeGraph is a community fork of this server optimised for local development: lighter setup, no managed deployment dependency, same graph-based storage model. memory-mcp is the canonical reference implementation; mcp-knowledge-graph is the local-first derivative. For personal workstations and development environments, mcp-knowledge-graph is the more practical choice. memory-mcp is the appropriate baseline for teams wanting the officially-maintained version.
Complements Context7 — context7 injects current library docs at query time; memory-mcp persists cross-session agent state. Relevant to BuildingEffectiveAgents and LLMSystemPatterns as the memory layer in agentic architectures. The graph structure (entities + relations + observations) aligns with the knowledge representation assumed by McpKnowledgeGraph.
Placed in Tools / Assess / inner. The official reference implementation for MCP-based agent memory; directly applicable to any team building agents with Claude or other MCP-compatible models. Inner position reflects immediate applicability and zero-friction adoption (part of the official MCP servers repo). Trial gate: agent retaining and correctly recalling at least five distinct facts across three separate sessions with no manual re-entry.
