Goal
Make contradiction and supersession handling a first-class, evidence-preserving memory capability: retain original memories, derive optional structured fact candidates, connect conflicts/supersessions, and use those signals only to rerank returned memories.
Why now
The current Korean retrieval contract exposes a real failure mode: a generic RRF correction improved relevance but also increased forbidden/contradictory results (baseline: hit@1 6.3%, hit@5 56.3%, MRR .226; forbidden@1 25.0%, forbidden@5 43.8%). The candidate was reverted. This issue addresses the root cause rather than tuning another rank weight.
Constraints
- Preserve raw source memories and their IDs; never overwrite/delete a prior memory merely because a later fact conflicts.
- No natural-language response policy, abstention policy, or language-specific ranking constants.
- Korean is the first validation language, but data model and retrieval policy must remain language-neutral with explicit extraction uncertainty/fallback.
- Keep LocalAdapter behavior compatible unless an explicitly versioned metadata field or API extension is required.
- CPU benchmark evidence only; GPU1 belongs to a different session.
Acceptance criteria
Initial status
Investigation and plan are in progress. Each phase will be posted as an issue comment with evidence and blockers.
Goal
Make contradiction and supersession handling a first-class, evidence-preserving memory capability: retain original memories, derive optional structured fact candidates, connect conflicts/supersessions, and use those signals only to rerank returned memories.
Why now
The current Korean retrieval contract exposes a real failure mode: a generic RRF correction improved relevance but also increased forbidden/contradictory results (baseline: hit@1 6.3%, hit@5 56.3%, MRR .226; forbidden@1 25.0%, forbidden@5 43.8%). The candidate was reverted. This issue addresses the root cause rather than tuning another rank weight.
Constraints
Acceptance criteria
Initial status
Investigation and plan are in progress. Each phase will be posted as an issue comment with evidence and blockers.