Skip to main content
An Agent with a memory space attached can remember — write a note during one run and recall it in a later one. Nora handles the mechanics: what to write, when to search, how to rank. You control the policy.

​
What a note looks like

Every memory item has:
  • Content — the text.
  • Type — free note, fact, event, procedure.
  • Importance — 0-1, defaults 0.5. Affects retention and retrieval boost.
  • Metadata — arbitrary key/value (entity ID, source URL, timestamp).
  • Scope — see Scoping.
  • TTL — optional expiration.

​
When the Agent writes

Writing is opt-in. Three patterns:

​
Automatic write on high-signal events

Turn on Auto-remember in the memory space attachment. The Agent decides when something is worth remembering — user provided a new preference, tool returned a stable fact, procedural step confirmed as effective. Writes are additive; nothing overwrites without explicit intent.

​
Explicit remember tool

Attach the remember tool to the Agent. The Agent calls it when it decides — a bit like a human choosing to write a note. Gives you visibility (every write shows in the Trace as a tool call) and control (you can tune the tool’s description to encourage or discourage writing).

​
Programmatic writes

From code or the CLI:
Useful for seeding a space with known facts, or importing from an existing store.

​
When the Agent reads

Reads are automatic. Every LLM call attached to a memory space triggers a retrieval — similar to knowledge retrieval, but scoped and with memory-specific ranking. Memory retrieval blends:
  • Semantic similarity to the current query.
  • Importance (higher-importance items ranked higher).
  • Recency (recent items get a modest boost, decays over time).
  • Access history (items retrieved and used effectively before rank higher).
Turn off recency boost for domains where old memory is as good as new (glossaries, established procedures).

​
Editing notes

Wiki spaces are directly editable in the UI. Click a note to open the editor. Edits update the embedding automatically. Index spaces are read/write from the Agent; not directly UI-editable (they’re often too high-volume). Delete or supersede specific items via the CLI or an admin action.

​
Deduplication

Two writes with substantially similar content within the same scope are deduplicated — the second becomes an update to the first (importance and timestamps refresh, content merges if slightly different). Prevents an Agent from writing “Alice prefers email” five times. Dedup threshold is adjustable per space (default: 0.9 cosine).

​
Consolidation

Periodically, Nora consolidates a space:
  • Related notes merge into stronger unified notes.
  • Low-importance items decay out.
  • Duplicate patterns collapse.
Runs automatically on a schedule (weekly by default) or triggered manually. Consolidation preserves the audit log — no information is lost, just reorganized.

​
Watching writes

The Trace view shows every memory read and write per run:
  • Reads: query, top-N items returned, scope.
  • Writes: content, importance, scope, whether it was a new item or a dedup update.
Great for verifying memory behavior in Draft before shipping.