Core
Memory
What the agent remembers about your organization, teams and sessions, and how to correct, share and control it.
Memory is what the agent has learned from your documents, answers and sessions. It is shown as a graph of entities and facts, and the agent uses it to answer with your context. You can read it, correct it, share it and switch it off.
Screenshot
Where to find it
Open the Memory page or Settings, Memory. Three places are available: Graph, Text and Settings. At the top you pick a scope to browse. When memory is still being built, the page says "Your memory graph is being updated. It appears here as soon as it is ready." An empty memory says "Nothing in memory yet."
Scopes
Memory is held at four levels, and you only see what you may read:
| Scope | Who holds it |
|---|---|
| Organisation | everyone in the organization |
| Team | the members of a team |
| User | you alone ("You") |
| Session | one conversation |
A team you are not a member of is not shown ("You are not a member of this team, so its memory is not shown.").
In short
The graph
The graph shows entities and the relations between them. Pick a Representation: Network, Galaxy, By scope, Table, Hierarchy, Timeline, Matrix or Radial, and Colour by type or scope. Search or filter ("Search or filter the graph") with words or filters such as type:person or scope:team:Alpha, narrow by type, scope, kind or time, zoom, and use Fit to view. Select a node to see Details and Graph about this. Export gives a PNG, SVG or JSON file.
You can correct what the graph holds: Rename, Change type, Add relation, Remove relation, Merge and Delete, each with Undo. Shared memory from another scope is read only here ("Shared memory is read only here").
Try it
Open the Graph place and search for the name of a customer or a vendor.
Select the node and read its relations in Details.
Open Graph about this to see only what relates to it.
Good to know
- A banner such as "Incomplete graph" tells you when a read could not serve some rows.
The text view
The Text place lists the facts as a table with search and paging. Select rows to apply an action to many at once: Pin, Unpin, Mark wrong, Mark right, Move to organisation, Move to team, Move to you and Forget. The app first counts how many rows would change, how many already are so, and how many are not yours, then asks you to Apply. Afterwards you can Undo.
Marks show where a fact stands: Stated (you said it), Inferred (the agent derived it), Pending, Overridden, Superseded and Marked wrong.
Good to know
- Past actions lists what was changed, or "Nothing done yet".
- "Nothing was removed: a saved chat still holds this text" means the text also lives in a stored session.
Promote and revoke
A session or a team memory can be promoted to a wider scope: a session to its team or to the organisation, a team to the organisation. The list Promoted shows what is shared, "Since" a date. Revoke takes it back. "Only whoever promoted it, or a team admin for the team, can revoke this."
In short
Settings
The Settings place has the controls for Use memory (switch retrieval on or off), Session capture (whether sessions are added to memory), the Retrieval mode, Hits per retrieval and the Default scope. Groups: Retrieval, Ranking, Lifecycle and Capture. Each knob shows its default and can be reset (Reset all, Undo the last change). Some values affect the whole deployment and need a second click ("This affects the whole deployment, not just this organisation. Click again to apply."). Only an administrator can change a value.
Good to know
- Reading memory needs the right to list it, changing the knobs needs the right to manage it. See Sign in and roles.
Use a document in answers
In the library, a document has the switch Use in answers (Do not use in answers turns it off). The confirmation says it is "a preference, not erasure under the GDPR: if the document is edited later, its new content is used in answers again." See Library and files.