Skip to main content
Wiki memory is a markdown notebook shared by people and the Agent. Like a team wiki: long-lived, edited in-place, and retrieved by semantic search.

​
When to use it

  • Standing operating procedures: “always look up the order status before starting a refund”
  • Team knowledge that isn’t in any external doc: tone, exception-handling patterns, tribal decisions
  • Findings the Agent needs to keep referencing across runs

​
Creating one

Memory → New space → Wiki (notes) creates one immediately. See Memory overview → Creating a space for the full flow.

​
The detail page

Opening a Wiki shows:
  • Notes tab: note list on the left, markdown editor on the right. A filter at the top narrows by scope.
  • Log tab: every event in chronological order (see below).
  • Settings tab: covered below.
The editor has three fields:
  • Path optional. The address the Agent references the note by. Rendered in monospace; use a folder-style path like docs/refund-policy to group related notes.
  • Title optional. Human-facing name.
  • Body markdown. An edit/preview toggle at the top of the editor lets you check the rendered output.
Saving refreshes the retrieval embedding automatically.

​
What you configure in Settings

Basics rename the space. Paradigm (Wiki) is fixed at creation. Flow binding bind the space to a specific Flow and every Agent in that Flow reads/writes it automatically. Leave unbound and you attach it per-Agent in the flow builder. Scope which request parameters partition notes. Combine flow_id / thread_id / user_id / custom keys. See Scoping. Access (optional) a partition key (auto-filter by caller identity) and a required-permissions gate to lock access at the space level. See Scoping → Access control. Memory agent (curator) the agent that manages this Wiki. Decides what to keep, grow, and retire.
  • Model runs extraction and consolidation. Pick from providers with keys registered in the workspace. Leave blank for a low-cost default.
  • Policy prompt free-text instructions for what to save vs discard. Example: “Keep stable customer preferences and decisions; discard chitchat; if a preference changes, supersede the old one.”
  • Multimodal how images, audio, and video reach memory (reuse / reference / re-analyze). Re-analyze costs extra.
  • Background writes when on, saves and deletes happen off the response path so latency isn’t affected. Reads are always synchronous. Recommended on.

​
Attaching to an Agent

In the flow builder, open the Agent, go to the Memory tab, and click Add memory. Cap runaway cost with Max memory cost on the same tab. If the whole Flow should share one Wiki, use Flow binding in the space’s Settings instead.

​
How notes get written

Three paths: The memory agent, automatically at the end of a run, the curator picks what to keep per the policy prompt above. Most notes land this way with no intervention. The Agent explicitly instruct the Agent in its system prompt to save findings. Shows up in the trace as a tool call, gives you tight prompt-level control over what gets saved. You directly, or via CLI write notes by hand from the Wiki detail page, or script it with nora memory. Handy for pre-seeding known facts or migrating from an existing store.

​
How notes get retrieved

Every model call from an Agent with the Wiki attached triggers a retrieval automatically. Similar to knowledge search, but scoped and with memory-tuned ranking.
  • Relevance semantic distance between the query and the note
  • Recency recent notes rank higher, then decay over time
  • Importance high-importance notes float up
Tune retrieval per-Agent: in the Agent Inspector’s Memory tab, each attached space has a Retrieval preset dropdown.

​
Rule-style notes

For standing rules, write notes that spell out the trigger and the action clearly: the memory agent will surface them well.
Add “always include rule-style notes in context for related turns” to the policy prompt and rules will land in the prompt every turn. There’s no built-in “rule” type: it’s just a convention for markdown notes.

​
Note maintenance

  • Edit click a note → edit in the editor. Save refreshes the embedding.
  • Exclude from search the ⋯ menu at the top-right of the editor → Exclude from search. Soft-deletes the note: it disappears from retrieval and a brief Undo toast lets you revert.
  • Dedupe near-duplicate notes in the same scope are merged automatically. The Agent can’t rewrite “Alice prefers email” five times.
See Forgetting for the full clean-up story.

​
Watching writes in the trace

Draft runs show every memory read and write in the trace view.
  • Reads: query, top-N items returned, scope
  • Writes: content, scope, whether it was a new item or a dedupe update
Handy for checking memory behaviour on a Draft before publishing.