Recall
There are four ways to get something back out of a memory. An agent opens every turn with the first; the last two are for when an answer must be checked.
briefing()
What this moment needs, for your system prompt: where things stand, what is due or was promised, the rules that apply, what happened, then the facts, each dated. Read, not generated: fast enough for every turn of a chat or a call.
search()
The memories that best match, ranked; search_full() adds the lessons, stories and open promises that bear on the query. No model call.
ask()
An answer in words from the memory only, with the memories behind it, or a plain message when nothing supports one.
context()
The memories an answer stands on, as a block for your own prompt, each with where it came from. When nothing bears on the question it says so.
When nothing is known
What context and ask return when nothing bears on the question, and why there is no setting to tune.
Which one to use
| Call | What you get | Model call | Typical time | Use it |
|---|---|---|---|---|
briefing(cue=...) |
A block for your system prompt: state, what is due, rules, what happened, facts | No | A fraction of a second | Every turn of a chat, every call |
search(q) |
The memories that best match, ranked; search_full adds lessons, stories and promises |
No | A fraction of a second | Looking something up mid-task |
ask(question) |
Our answer in words, and the memories behind it | Yes | A couple of seconds | When an answer must be checked |
context(question) |
The memories an answer stands on, as a prompt block, each with its source | Yes | A couple of seconds | When your own model writes the answer and nothing may be invented |
A voice agent should never wait on a model call before it speaks: give it the briefing. ask and context have a
model judge whether anything supports an answer, which is what keeps them from inventing, and what takes the time.
All four keep to labels where they take them, to read only what the sources carrying them said. See
Labels.