Skip to main content
Recall is the first step of the memory loop. Before the agent responds, you ask the memory system for context relevant to the upcoming message. It returns a single prompt string — ranked, deduplicated, and packed to a character budget — ready to drop into the agent’s system prompt.

Assemble context

The returned string is the only thing the model sees. You don’t handle raw database rows, ranking, or deduplication — the memory system does that internally and hands you a compact block.

Parameters

Injecting into the agent

Pass the context as part of the system prompt when you start the stream:
The agent now answers with the prior conversation and stored facts in mind, without you manually loading history.

Scope rules

Recall respects the same scope as the rest of memory: the agent + user pair. Working context and transcripts are thread-scoped, so they only appear when threadId matches. Durable facts cross thread boundaries, so the agent still remembers pinned preferences even in a brand-new thread.