Memory
Give agents a long-term memory store so they recall what matters in later threads.
By default an agent starts every thread fresh: it knows its instructions, its skills and what is in the conversation, nothing more. A memory store gives it a long-term memory. Attached to an agent, the store lets the agent save durable facts, such as a client's reporting preferences or how a team names its share classes, and recall them in later threads.
A store holds memories, one per subject. Several agents can share one store, which is how a team's knowledge is shared deliberately rather than by default.
Open the Memory page
Memory is part of Agent Studio. Open the Settings tab and choose Memory in the Agent Studio group, or click the Memory card on the Ecosystem page. The page, Memory stores, says what stores are for: Where agents save memories to recall in later threads.
It lists the workspace's stores, each with its name, what it is for and whether it is Active. A lock after a name marks a store that is human-only: hovering it says Human-only: agents read it but cannot update it. Use the search to filter, and the archive button to show archived stores, which stay in the list greyed out.
Create a memory store
- On the Memory stores page, click New.
- In New memory store (Agents bound to the same store share what it remembers.), give it a Name and a Description: what the store is for, and which agents should share it.
- Leave Agents can update on to let agents save to the store. Turn it off to make the store human-only (Off makes the store human-only: agents read it but cannot save to it.).
- Click Create.
A store's ⋯ menu opens it (Open memory store), opens its Settings, activates or deactivates it, archives or restores it, and deletes it. Deleting a store deletes its memories and leaves the agents bound to it with none.
Attach memory to an agent
- Open the agent in Agent Studio › Agents.
- In the Memory section, choose a store. No memory leaves the agent without one.
- Decide whether the agent saves on its own with Remember automatically:
- Off: when something worth keeping comes up, the agent answers, then offers in one sentence to remember it. It saves only after you say yes.
- On: the agent saves what is worth keeping once the rest of its work is done, and ends its answer by saying what it saved.
On a human-only store, Remember automatically is disabled, and hovering it explains that agents only read that store.
The section's link opens the bound store (Manage this memory store) or the Memory page.
Read and edit memories
Open a store to see it on a page of its own. Its header shows the store's name and state, with a Human-only badge when agents can't save to it, and its description; at the right, badges say how many memories it holds, how many agents use it, when and by whom it was created, and its ID. Below, its memories sit in two panes.
The left pane lists the memories under their tags, with untagged ones first; a memory with several tags shows under each of them. Search them, fold every tag with Collapse all, show archived memories, or click + (New memory). The right pane shows the selected memory:
- Its subject, and a badge saying who wrote it: Written by an agent, Written by a person or Imported.
- Its facts: Tags, when it was Updated, and its Summary.
- Its Content, in a card of its own.
From the memory's ⋯ you can Edit, Archive (or restore) and Delete it.
Write a memory yourself
Click + in a store's memory list. In New memory, fill in:
| Field | What it holds |
|---|---|
| Subject | What the memory is about, for example "Home and location". |
| Summary | What the subject covers, in a line. Agents see this to decide whether to load the memory. |
| Tags | Comma-separated labels, such as "client, reporting", used to group memories. |
| Content | Everything worth knowing about the subject, loaded only when the agent asks for it. |
Click Create. Subject takes up to 60 characters and Summary up to 80, each shown as a live count under the field.