Choose memory
Open Blueprint → Memory. Turn on the kinds of memory the agent needs, then test a conversation long enough to use them.
Recent conversation
Section titled “Recent conversation”Recent conversation keeps recent messages in the agent’s context. It helps the agent answer follow-up questions, such as recalling an order code supplied earlier in the chat.
Open Recent-message limit to choose how many messages to include. More messages increase the context sent on each reply. Leave the number blank to use your organization’s default. Set it to zero to disable recent-message history.
Long-term conversation memory
Section titled “Long-term conversation memory”Long-term conversation memory summarizes older messages. It uses extra model calls as the conversation grows.
Open Advanced settings and keep Scope set to Per Thread, for one conversation. Leave the model fields blank to use the agent’s configured model. To choose another model for memory, enter an explicit model ID from the model catalog in Shared Model (observer + reflector). Use a model your organization can run.
The observer summarizes messages. The reflector condenses those summaries. A shared model runs both. Observer Model and Reflector Model override it for their respective calls. A blank field uses the shared model, or the agent’s configured model when no shared model is set.
Keep the token thresholds unchanged to start. A short exchange can succeed without reaching the threshold that triggers a summary. Test a longer conversation and ask the agent to recall details from its beginning.
Saved facts and task details
Section titled “Saved facts and task details”Saved facts and task details keeps selected facts in a template. Open Saved-facts settings to describe the information to keep, such as a customer’s order code or the current task.
Choose who updates those facts. When long-term conversation memory is also on, the background observer updates them.
Use fictional details for tests. Ask the agent to remember a detail, then ask for it in a later turn. Check the actual reply before relying on the setup.
Limits
Section titled “Limits”These defaults come from the limit catalog. Your organization’s settings can differ.
| Limit | Default | Support can change it |
|---|---|---|
| Most recent messages an agent may load into context on each turn | 200 | Yes |
| Recent messages loaded into context when the blueprint does not set a number | 10 | Yes |
| Most characters of MEMORY.md injected into the agent’s context each step | 16,000 | Yes |