[Feature] Allow Batch/On-Demand Memory Extraction Instead of Per-Message Processing

Is your feature request related to a problem?

Currently Memory Agent processes every conversation turn immediately. In my use case, this causes unnecessary overhead: real-time extraction isn't critical, and many turns are low-value (greetings, clarifications). Extracting every single message wastes tokens when only 3-4 turns in a 20-turn session contain meaningful user information.

The Solution

Allow memory extraction to run on-demand or in batch mode at conversation boundaries, rather than automatically after every message.

1. Trigger extraction manually or on session end
2. Keep current per-message behavior as default, add batch mode as optional setting

Alternatives Considered

No response

Affected Components

No response

Participation

• I am willing to submit a pull request for this feature.