Which memory to read, and how far back to reach in it.
Recall decides how far back to reach. The paths themselves come from the memory: this component reads the namespaces its strategies file records under, so what you configure in AWS is what it looks in.
A namespace that is not tied to a person is never read.
Recent is this conversation's own messages. Long-term is what AgentCore extracted from past conversations.
How much memory to return, and what to search for.
Returning {{ (config.short_term_limit * 1) || 12 }} recent messages and {{ (config.long_term_top_k * 1) || 5 }} long-term memories, found by {{ config.long_term_query_mode == 'preferences_profile' ? 'asking for standing preferences' : 'matching the message' }}. Tick Advanced settings to change any of it.
The counts decide how much reaches the model. There is no separate character budget to keep in step with them.
Standing preferences is the mode that answers "what do you know about me" - a question that names nothing for the search to match on.