Memory MCP

A knowledge-graph-based memory system for AI agents that enables persistent information storage between conversations.

Features

• Persistent memory storage using a knowledge graph structure

• Entity-relation model for organizing information

• Tools for adding, searching, and retrieving memories

Related MCP server: cortex-engine

Tools

The system provides the following MCP tools:

• load_knowledge_graph(): Retrieves the entire knowledge graph

• get_knowledge_graph_size(): Returns the current size category of the graph ("small", "medium", or "large")

• add_entities(entities): Adds new entities to the memory

• add_relations(relations): Creates relationships between entities

• add_observations(entity_name, observations): Adds observations to existing entities

• delete_entities(entity_names): Removes entities from memory

• delete_relations(relations): Removes relationships

• search_nodes(query, search_mode): Searches for entities and relations matching a query. Supports three search modes:

  • "exact_phrase": Matches the entire query as a substring

  • "any_token": Matches if any word in the query matches (default)

  • "all_tokens": Matches if all words in the query match

• open_nodes(names): Retrieves specific entities and their relationships between them

Usage

Run the agent with:

uv run memory_agent.py

The agent will automatically:

1. Load its memory at the start of conversations

2. Reference relevant information during interactions

3. Update its memory with new information when the conversation ends

Exit a conversation by typing q.

Configuration

Set the memory storage location with the MEMORY_FILE_PATH environment variable (defaults to memory.json).