docs: discuss DeepRefine and long-term agent memory systems

[Pillowman777]

Summary

• clarify the distinct research focus and triggering mechanisms of DeepRefine and long-term agent memory systems
• describe defect localization, structured refinement, and downstream validation as complementary to acquisition, organization, and retrieval
• outline concrete potential extensions inspired by systems such as GBrain, including automatic triggers, freshness signals, contradiction handling, provenance tracking, and scheduled refinement

Scope

Documentation-only change. The roadmap items are explicitly presented as potential extensions, not implemented features.