[Feature]: Support skipping memory extraction for scheduled / ephemeral tasks (e.g. daily cron jobs)

Issue Origin

Observed or reproduced in a real environment

Problem Statement

In real-world AI agent deployments, agents frequently run automated scheduled/cron tasks (for example, daily news search and push, RSS monitoring, disk health check, weather forecast).

Currently, OpenViking executes memory extraction on sessions indiscriminately upon commit/session end. For periodic tasks like daily news search, each run may fetch and generate dozens or hundreds of kilobytes of transient information. Automatically extracting memories from these ephemeral tasks leads to several issues:

1. Memory Pollution: Ephemeral daily news and temporary execution logs pollute the user's long-term memory store.
2. Token Waste: Large volumes of transient web search / article content are sent to the LLM for memory distillation without long-term value.
3. Risk of Upstream LLM Filter Triggers: News articles frequently contain sensitive geopolitical keywords or real-time news terminology that can trigger upstream LLM content moderation filters (e.g., MiniMax sensitive input filters), resulting in API errors.

Proposed Solution

We would like to request an official switch or mechanism to skip/disable memory extraction for scheduled or ephemeral sessions:

• Allow specifying whether a session should undergo memory extraction (e.g., via session creation options, commit options, or session-level configuration/prefix patterns).
• When extraction is disabled for a session, the agent can still freely use context and retrieval tools during the task, but OpenViking will skip long-term memory extraction when the session finishes.

Alternatives Considered

Currently having to maintain custom interceptors in upstream agent plugins or manually prevent sessions from being committed, which is fragile across container updates.

Feature Area

Session Management

Use Case

Agent platforms running recurring cron jobs (daily news briefings, system health checks, notifications) where the agent needs context/tools to perform the job, but the task's conversational transcript must not be extracted into permanent user memory.