Skip to main content
A memory space is a named container with a fixed type. You create spaces in the workspace, then attach them to Agents.

​
The four types, in detail

​
Wiki

Human- and Agent-editable notes, retrievable by semantic search. Best for:
  • Standing operating procedures.
  • Team knowledge that isn’t in any external doc.
  • Playbooks the Agent should follow.
Notes have a title, body, tags. Editable in the UI. The Agent can also write new notes when granted permission (see Notes & recall). Retention: indefinite unless you configure a TTL.

​
Index

A vector store the Agent writes to and reads from. Best for:
  • Facts about entities (customers, products, incidents).
  • Anything you’d think of as “records” that don’t fit the Wiki text-note model.
  • Ingested transcripts, emails, chat logs the Agent references.
Each item has a content field (embedded), metadata, and importance score. Retrieval uses the same hybrid search as knowledge collections. Retention: importance-decayed by default. Highly-important items persist longer; low-importance items decay out.

​
Session

Automatic per-conversation memory. The Agent doesn’t do anything explicit — every chat message is captured, summarized as the conversation grows past the model’s context window, and made searchable. Best for:
  • Chat Agents that need “we talked about X earlier” behavior.
  • Multi-turn tasks where earlier turns matter later.
Retention: per-session unless promoted. You can pin an important turn to persist across sessions.

​
Evolution

Read-mostly space that auto-populates with patterns from successful improvements. When an improvement to a similar situation was validated and shipped, the pattern lands here so future runs can benefit. You don’t write to Evolution directly — it grows through the improvement workflow. But Agents attached to it get the benefit automatically. Best for:
  • Long-running Agents where you want compounding quality.
  • High-volume Agents where the improvement flow generates rich patterns.
Retention: patterns persist while they remain useful (evidenced by re-use). Unused patterns decay after 90 days.

​
Creating a space

Workspace → Memory → New space. Configure:
  • Name — human-readable.
  • Type — Wiki / Index / Session / Evolution.
  • Default TTL — for items without an explicit TTL.
  • Retention policy — hard delete vs. supersede.
  • Access — which Agents can read, which can write.

​
Attaching to an Agent

Agent Inspector → Memory → Add space. Pick from workspace spaces. Choose read-only or read/write.

​
Naming spaces

Suggested pattern:
  • <Agent role>-<content> — e.g., support-customer-history, sales-lead-notes.
  • Or <domain>-<type> — e.g., billing-facts, product-glossary.
Consistency helps when you have many spaces.

​
Space size and cost

  • Wiki / Session: cheap. No embedding cost per item.
  • Index: has embedding cost — same rate as document ingest.
  • Evolution: managed by Nora; no direct cost.
Storage is included up to plan limits; overage is priced per GB/month.