3.1 Memories

A Memory defines which data an Agent can reach when it answers a query. It is the bridge between an agent and a data source: you create a Memory over some data, initialize it, and then attach it to an agent on the Agent System canvas. The same Memory can be attached to several agents, and one agent can hold several memories.

There are three types of Memory:

• Graph: Covers Target Entities from the knowledge graph in the Data Context Hub. You select the entities to include, and the agent retrieves structured data and relationships from them. A Graph Memory uses an Embedding Configuration to vectorize its content for similarity-based retrieval.
• Projections: Covers curated Views and Functions from the Projections section, so an agent can work on prepared projections of the data rather than the raw graph.
• File: Covers documents ingested by a File Intake Agent. You do not create a File Memory here (the Source type selector offers only Graph and Projections); the File Intake Agent creates and maintains it.

You create Graph and Projections memories from the Memories section. Give a new memory a title and description, choose its Source type, configure its sources, then Save.

Graph Memory

A Graph Memory exposes selected Target Entities to an agent.

1. In the Memories section, click Add new, enter a title and description, and set Source type to Graph.
2. Choose an Embedding configuration from the dropdown (or leave it on Default). This is the Embedding Configuration used to vectorize the selected content.
3. Under Select target entities, search for and move entities from Available to Selected. Use the + next to an entity to also add its related entities, remove a single entity with its x, or Remove all to clear the selection.
4. Click Save.

Initialize

A saved Memory still has to load its data before an agent can use it. Use the load action in the Memory's Actions section:

• The button reads Initialize the first time (and again if a load previously failed, was canceled, or the data is stale), and Load when it just needs a refresh.
• While it runs, a live percentage and a spinner show progress; hover the spinner to cancel the load.
• The overview list shows each Memory's Load status (Queued, Processing, Finished, Error, or Canceled) and Last loaded at.

caution

The embedding model fixes the vector dimensions for a Memory. If you change the Embedding configuration after initializing, re-initialize the Memory so its vectors are regenerated with the new model.

Projections Memory

A Projections Memory exposes selected Views and Functions from the Projections section.

1. Click Add new, enter a title, and set Source type to Projections.
2. Under Select published views and Select published functions, tick the items to include. Each list has its own search box, and your current selection is shown as chips below it.
3. Click Save.

A Projections Memory does not use an embedding configuration.

File Memory

A File Memory holds documents (for example PDFs) uploaded through a File Intake Agent. The Intake Agent chunks each file's contents, stores the chunks so they can be retrieved, and creates the Memory and its entities automatically. You open an existing File Memory from the overview list to inspect it, but you do not create one from the Memories page, and deletion is handled through the Intake Agent, not here.

Delete a Memory

Open a Graph or Projections Memory and click Delete memory, then confirm. File Memories are removed through the File Intake Agent.