Central-Memory-MCP

Architecture Guide (.NET Implementation)

Overview

The Central Memory MCP Server is a serverless Azure Functions (.NET 10 isolated worker) application exposing Model Context Protocol (MCP) memory & knowledge graph operations. It stores entities, relations, observations, and statistics in Azure Table Storage with strong workspace isolation.

Serverless Azure Functions (.NET 10 isolated) implementing a minimal MCP memory graph (alpha). Current scope: entities + relations with basic CRUD and graph read. Workspace isolation via WorkspaceName partition key.

High-Level Diagram

flowchart LR
    C[Client / MCP Tool] -->|read_graph / upsert_entity / upsert_relation / get_entity_relations| F[GraphFunctions]
    F --> S[KnowledgeGraphService]
    F --> R[RelationService]
    S --> T1[entities Table]
    R --> T2[relations Table]

Implemented Components

1. GraphFunctions - MCP tool bindings (4 tools implemented).
2. HealthFunctions - /api/health, /api/ready.
3. KnowledgeGraphService - entity upsert, single entity lookup, full workspace read.
4. RelationService - relation upsert, relations-from-entity lookup, workspace relations list.
5. TableStorageService - ensures existence of tables and returns clients.
6. Models - EntityModel, RelationModel, WorkspaceModel (workspace not yet used in tool surface).

Data Model

public record EntityModel(string WorkspaceName, string Name, string EntityType, List<string> Observations, string? Metadata)
// PartitionKey = WorkspaceName, RowKey = Guid (Id)
// Observations persisted joined by "||" delimiter

public record RelationModel(string WorkspaceName, Guid FromEntityId, Guid ToEntityId, string RelationType, string? Metadata)
// PartitionKey = WorkspaceName, RowKey = Guid (Id)

Storage

• Tables: entities, relations, (workspaces reserved)
• Partition strategy: simple partition per workspace for both tables.
• Row keys: GUID for stability (names may contain special chars).

Operations Flow (Entity Upsert)

sequenceDiagram
    participant Tool as MCP Tool
    participant Func as GraphFunctions.UpsertEntity
    participant Svc as KnowledgeGraphService
    participant Tbl as Azure Table (entities)
    Tool->>Func: UpsertEntityRequest
    Func->>Svc: UpsertEntityAsync(model)
    Svc->>Svc: Lookup existing by name
    Svc->>Tbl: UpsertEntity
    Tbl-->>Svc: Success
    Svc-->>Func: EntityModel (Id preserved/reused)
    Func-->>Tool: { success, id, name }

Error Handling

• Request validation (required fields checked). Returns { success=false, message="..." }.
• Relation upsert fails if source/target not resolvable.
• Missing entity in get_entity_relations returns friendly message.

Logging

(Planned) Structured logging with workspace scope; currently minimal.

Security

Development assumes trusted environment; add auth (Azure AD) before external exposure.

Roadmap

• Additional tools: search, stats, batch operations.
• Metadata merge & observation limit controls.
• Duplicate detection & merge.
• Vector similarity index.

Deployment

Publish via dotnet publish and Azure Functions tooling. Health endpoints for readiness probes.

This site is open source. Improve this page.