Data Model

The engine uses PostgreSQL (via Prisma) to store memory metadata and Qdrant for vector embeddings. The primary data entity is Memory.

Schema (`Memory`)

Field	Type	Description
`id`	`Int`	Unique primary key (autoincrement).
`userId`	`Int`	ID of the user who owns the memory.
`agentId`	`String?`	Optional ID of the agent that created the memory.
`runId`	`String?`	Optional ID for grouping memories by execution run.
`role`	`String?`	Role of the message creator (e.g., "user", "assistant").
`source`	`String`	Source of the memory (e.g., "chat", "codebase", "manual").
`sourceId`	`String`	External ID references the source (e.g., message ID).
`timestamp`	`DateTime`	Logical timestamp of the memory.
`content`	Stored in Qdrant	The actual text content is primarily managed via embeddings/payloads, though raw text can be retrieved.
`contentUrl`	`String`	URL or path to the original content.
`title`	`String`	Title or brief description.
`origin`	`String`	Origin system or component.
`tags`	`String[]`	Array of tags for categorization.
`category`	`String[]`	Array of categories.
`attribute`	`Json`	Flexible JSON field for extra metadata (e.g., key/value pairs).
`summary`	`String`	Short summary of the memory.
`type`	`String`	Type of memory (e.g., "text", "procedural").
`importance`	`Float`	Importance score (0-1).
`confidence`	`Float`	Confidence score (0-1).
`embeddingRef`	`Int`	Reference ID potentially linking to vector store status.
`contentHash`	`String`	Unique hash of the content for deduplication.
`createdAt`	`DateTime`	Creation timestamp.
`updatedAt`	`DateTime`	Last update timestamp.

To ensure performance, the following indices are maintained: