Memory

Memory is nara’s knowledge store. It holds typed objects — structured records whose shape is defined by a data type — that agents can store, retrieve, and search during conversations and workflows. Objects can be linked with relationships, forming a knowledge graph that agents traverse to find related information.

How agents use memory

Agents interact with memory through a set of knowledge-graph functions available as tools:

• Query Memory — semantic search across stored objects, used to answer questions from organizational knowledge.
• Neighborhood Search — search within the graph neighborhood of a given object to find closely related information.
• Store Object, Retrieve Object, List Objects — create, fetch, and enumerate typed objects.
• Get Node and List Neighbors — inspect a single object and the objects linked to it.
• Create Relationship, Delete Relationship, List Relationship Types — manage the links between objects.

When an agent uses these functions in chat, you see status phrases such as “Searching memory” while the call runs. Which functions an agent may use depends on its tool assignment in the agent editor.

Memory Explorer

Open Memory Explorer at /admin/explorer (admin sidebar, under Knowledge) to organize memory as a folder tree.

• Create folders to group objects by topic, team, or source.
• Drag and drop objects and folders to move them within the tree.
• System-managed nodes are locked and cannot be moved or deleted — nara maintains these itself.
• Delete empty folders, or delete individual objects from the tree.

The Explorer also shows a type overview, so you can see which data types the objects in a folder use.

Browsing and editing objects

Objects can be browsed in two places:

• Admin views — /admin/objects/type/[type] lists all objects of a data type; /admin/objects/[id] opens a single object.
• Objects page — the user-facing /memory route (nav label Objects) lets regular users browse memory without admin access.

Both list views offer Cards and Table layouts with pagination. Opening an object shows its detail view, where the contents are rendered according to the object’s data type — each field appears with its defined label and type. From the detail view you can edit individual fields or delete the object.

Ingestion Deck

The Ingestion Deck turns documents into memory objects. Open it with the Ingest button.

1. Upload files. Drag and drop documents onto the deck. Supported formats: PDF, DOCX, HTML, TXT, CSV, and images. Each file shows its processing cost in credits.

2. Watch the pipeline. Each file moves through four phases: Intake → Extracting → Reviewing → Committing. During extraction, nara reads the document and proposes structured objects that match your data types.

3. Review drafts. Before anything is written to memory, extracted objects appear as drafts for review. Draft notifications also appear in Notifications, so you can return to a pending review later.

4. Commit. Approve the drafts you want to keep. Committed objects become regular memory objects — searchable by agents and visible in the Explorer and object views.

Review is the safeguard: nothing enters memory until you have checked the extracted objects, so a poorly extracted document cannot silently pollute your knowledge base.