Current Memory interface is too basic - only handles conversation messages. I think this most like session(Short-term memory), In agent development, I've found two common needs that the current Memory interface doesn't support:

1. Session State: Agents need to store temporary data within conversations (user preferences, current task, form data, some kv) beyond just messages.
2. Memory: Long-term, searchable knowledge across sessions

The current Memory interface only handles message storage, but production agents need kv state management for better user experience.
Additionally, Context Management should be a separate abstraction to make it easier for framework users to manage context - providing simple APIs to control what information gets included in prompts, handle context window limits, and organize contextual data without manually managing token counts and message filtering.