Memory

Implements a graph-based memory system that stores entities (people, concepts, events) and their relationships as persistent nodes and edges, enabling structured knowledge representation beyond flat key-value storage. The system uses a graph data model where entities are nodes and relationships are directed edges with semantic labels, allowing LLM clients to query and traverse connected knowledge through MCP tool calls. This approach enables contextual memory recall where related entities are discoverable through relationship traversal rather than keyword matching alone.

Provides MCP tools for creating and updating entities (discrete knowledge units) with configurable types and metadata fields, organizing memory around named entities rather than unstructured text. Each entity is a node with a type identifier (e.g., 'person', 'project', 'concept') and arbitrary metadata properties, stored in the graph structure. This enables type-aware queries and filtering where clients can retrieve all entities of a specific type or update entity properties without affecting the graph structure.

Implements directed graph edges between entities with semantic labels (e.g., 'worked_on', 'knows', 'depends_on'), enabling clients to define and query relationships that carry meaning beyond simple connections. Relationships are first-class objects with labels and directionality, allowing traversal queries like 'find all projects this person worked on' or 'find all people who know each other'. The system supports both creating new relationships and querying existing relationship paths through MCP tool calls.

Provides MCP tools for querying the memory graph using entity names, types, and relationship traversal patterns, returning structured results that include connected entities and their relationships. Queries can filter by entity type, search by name patterns, and traverse relationships to find connected entities, all exposed as discrete MCP tools. The system returns full entity records with metadata and relationship information, enabling clients to understand both the entity and its context in the graph.

Implements the Memory server as an MCP server that exposes all memory operations (entity creation, relationship management, queries) as callable tools through the Model Context Protocol, enabling LLM clients to invoke memory operations as part of their reasoning loop. The server uses MCP's tool registration mechanism to define tool schemas with input/output types, allowing clients to discover available memory operations and call them with structured parameters. This integration makes memory operations first-class capabilities available to any MCP-compatible client.

Stores all memory data in-process memory (JavaScript objects/maps) scoped to the server session, providing fast access and isolation between different client sessions but no persistence across server restarts. Each server instance maintains its own graph in memory, meaning memory is lost when the server stops and is not shared between concurrent clients unless explicitly synchronized. This design prioritizes simplicity and performance for reference implementation purposes over durability.

Defines MCP tool schemas for each memory operation (create entity, add relationship, query) with input parameter types, output types, and descriptions, enabling MCP clients to discover available memory operations and understand their signatures. The server registers these schemas with the MCP protocol, allowing clients to list available tools and understand what parameters each operation expects. This enables proper tool calling with type validation and helps clients understand the memory API surface.