Research and improve Memory Search algorithms

[Xunzhuo]

Goal

Research and improve Memory Search algorithms for higher-quality retrieval over Personal Model claims, Steps, and contextual recall inputs.

Scope

• Evaluate current lexical, fuzzy, CJK n-gram, semantic, and claim-confidence ranking behavior.
• Improve query expansion / query variants, cross-language recall, topic matching, stance/polarity handling, and no-match calibration.
• Distinguish claim-aware search from conversation search while allowing useful combined ranking signals.
• Add diagnostics that explain why a result matched, ranked, or failed.
• Preserve safety: weak matches must not become asserted Personal Model truth.

Acceptance Criteria

• Current memory search behavior is benchmarked with representative examples.
• Proposed algorithm changes include measurable quality targets.
• Golden tests cover strong match, weak match, no match, cross-language, CJK, and conflict-prone queries.
• Diagnostics make ranking decisions inspectable for development and evaluation.
• The algorithm remains aligned with docs/system-design/system-layer-model.md.