Central-Memory-MCP

Storage Guide (Current Implementation)

Overview

Alpha implementation persists entities and relations in Azure Table Storage. Workspace isolation via WorkspaceName (PartitionKey). Row keys are GUIDs (no reliance on names). Observations stored as a single delimited string (||).

Tables

• entities
  • PartitionKey: WorkspaceName
  • RowKey: Entity Guid (N format)

  • Properties: Id, WorkspaceName, Name, EntityType, Observations (“

    ” joined), Metadata

• relations
  • PartitionKey: WorkspaceName
  • RowKey: Relation Guid (N format)
  • Properties: Id, WorkspaceName, FromEntityId, ToEntityId, RelationType, Metadata
• workspaces (reserved for future use)

Example Entity Row

PartitionKey = demo
RowKey = 5f3a2e9d2c8f4a1e9d3b2c7a5e4f1d0a
Name = Alice
EntityType = Person
Observations = Software engineer||Loves Go
Metadata = {"dept":"Engineering"}
Id = 5f3a2e9d2c8f4a1e9d3b2c7a5e4f1d0a

Upsert Strategy

1. Lookup existing entity by PartitionKey + Name (OData filter)
2. If found, reuse existing Id
3. Replace entire row (Replace mode)

Relation Upsert

• Requires existing entity GUIDs or resolves names first
• Reuses relation Id if identical From/To/Type previously upserted (handled in service layer)

Observations Format

Stored joined by ||. On read split into list:

string.Join("||", observations)
// Read: value.Split("||", RemoveEmptyEntries)

Workspace Isolation

Single partition per workspace keeps queries simple; scalability acceptable for early stage. Future optimizations may introduce composite partitioning if hot partitions emerge.

Configuration

Local:

AzureWebJobsStorage=UseDevelopmentStorage=true

Production example:

AzureWebJobsStorage=DefaultEndpointsProtocol=https;AccountName=<name>;AccountKey=<key>;EndpointSuffix=core.windows.net

Use managed identity / Key Vault for secretless access (future enhancement).

Query Patterns

• Read graph: filter PartitionKey eq '<workspace>'
• Lookup by name: adds and Name eq '<escaped>'
• Relations from entity: filter by PartitionKey then client-side match (future: add FromEntityId indexed property query optimization)

Error Handling

• Invalid payloads produce { success:false, message }
• Missing referenced entity aborts relation upsert.

Limitations / Roadmap

• No pagination (entire workspace scan)
• No secondary indexes (e.g., by type)
• No search by observation content
• No batch writes beyond per-entity upsert

Future improvements: batch APIs, observation archival, indexed queries, search entities by prefix/type, relation reverse lookups.

This site is open source. Improve this page.