MWGMorningwood/Central-Memory-MCP

Model Context Protocol (MCP) compliant memory & knowledge graph server implemented in .NET 10 (Azure Functions isolated worker). Provides durable project memory (entities, relations, observations) for AI assistants with workspace isolation and simple HTTP tool endpoints.
Implemented MCP Tools
Current function set (alpha stage):
- read_graph - returns all entities for a workspace (relations currently queried separately and joined)
- upsert_entity - create or update entity (preserves existing Id if name exists)
- upsert_relation - create or update relation between two entities (requires GUIDs or resolvable names)
- get_entity_relations - list relations originating from a specific entity
- Health & Ready endpoints (/api/health, /api/ready)
Planned (not yet implemented): search_entities, search_relations, stats, temporal, batch operations, merge/detect duplicates.
Data Model
- Entities table: PartitionKey = WorkspaceName, RowKey = Guid (Id)
- Relations table: PartitionKey = WorkspaceName, RowKey = Guid (Id)
- Workspaces table (future expansion)
Observations stored as a single delimited string ("||") internally; split into List at read time.
Quick Start
To read the graph, invoke the read_graph MCP tool from the client and supply the workspaceName parameter.
Directory Layout
Usage Notes
- Use workspaceName consistently; workspaceId in docs replaced.
- Upsert preserves entity identity by lookup on (WorkspaceName + Name).
- Relation upsert requires entity GUIDs or resolves names; fails if names missing.
Logging & Telemetry
Add Application Insights connection to capture request latency & storage dependency tracking (future enhancement).
Roadmap
Add search, pagination, stats aggregation, batch operations, duplicate detection, merge strategy, semantic vector layer.