Epic: Legacy memory enhancement via Neo4j retrieval reuse

[animaworks-dev-team]

Overview

Enhance the Legacy (ChromaDB + RAG) memory backend by extracting and reusing Neo4j hybrid retrieval components (RRF, cross-encoder rerank, confidence gate) while preserving Legacy recall strengths observed on LoCoMo (63.7% overall F1 vs Neo4j 55.5%).

Problem

• Legacy scope_all lacks cross-encoder rerank and retrieval confidence gating.
• Ingest uses coarse session chunks without atomic facts, entity index, or ingest-time invalidation.
• Priming and search_memory use divergent ranking paths.

Approach

Split into 10 child issues (Wave 1–3):

1. C1: Shared retrieval module extraction (core/memory/retrieval/)
2. C2: Legacy cross-encoder rerank integration
3. C3: Retrieval confidence gate (adversarial abstention)
4. C4: Atomic fact extraction + bi-temporal metadata
5. C5: Entity index + search-time boost
6. C6: Ingest-time contradiction / valid_until
7. C7: Priming ↔ search unification
8. C8: Rule-based query expansion
9. C9: Access-count LTP boost in rerank
10. C10: LoCoMo Legacy regression smoke harness

Acceptance Criteria

• Legacy LoCoMo F1 ≥ 63.7% baseline after Wave 1
• Legacy open_domain F1 ≥ 68%
• Legacy adversarial F1 ≥ 82% (target after C3)
• Neo4j backend tests remain green
• All 10 child issues completed

---

Full specification: docs/issues/20260522_legacy-memory-enhancement-epic.md (local)