Phase 3 — Goals

Track A — Memory Schema & Data Foundation

Goal 3.A: Establish the memory schema with HNSW, temporal validity, and multi-label taxonomy

Milestones:

• 3.1.1 — Postgres schema v2 (completed): expand foundation ibex_core.memories (from G5.M1 / ADR-0047) with HNSW vector index (m=16, ef_construction=64), quality columns, generated search_vector, bounded embedding_model/metadata CHECKs. Labels + primary sync already shipped in ADR-0048; expand ADR is ADR-0052 (not ADR-0040).

Acceptance criteria:

• Migration applies cleanly from zero via make db-migrate
• HNSW index type confirmed in CI assertion (not just present — type-checked)
• RLS integration test proves cross-org denial on memories (labels covered in 2.5.G5.M2)
• Primary-category sync verified via existing trigger (insert 2 labels, differing confidence, correct one wins)

---

Track B — Vector Store & Embedding Integration

Goal 3.B: Pluggable vector store with category-conditional composite scoring

Milestones:

• 3.2.1 — VectorStore interface + pgvector-HNSW implementation: VectorStore ABC, PgVectorStore with SET LOCAL hnsw.ef_search per-transaction. Composite scoring v2 with category-conditional half-life (factual: 180d, procedural: 120d, preference: 45d, behavioral: 30d, episodic: 14d). Embedding client calling Phase 2.5's embedder service. Recall/latency benchmark at 10K/100K/1M synthetic rows.

Acceptance criteria:

• VectorStore ABC with PgVectorStore implementation, ef_search set per-transaction
• Composite scoring v2 unit-tested against documented formula
• Benchmark run at 10K/100K/1M rows, results committed
• embedding_client.py has zero direct ML library imports

---

Track C — Memory Write Pipeline

Goal 3.C: 9-step write pipeline with PII detection, dedup, and temporal conflict resolution

Milestones:

• 3.C.1 — PII detection and redaction: Two-tier detection via Microsoft Presidio (regex patterns + NER). Redaction before embedding. Quarantine for low-confidence PII findings.
• 3.C.2 — Content and near-duplicate detection: SHA-256 content hash for exact dedup. Cosine > 0.92 for near-duplicate candidate detection.
• 3.C.3 — Temporal-interval-aware conflict detection: Interval overlap check before LLM escalation. Sequential facts (valid_from ordering) auto-classified as supersedes with zero LLM calls. ADR-0042 rewritten for temporal-first design.
• 3.C.4 — Multi-label memory classification: memory_labels join table. Extraction prompt outputs label array with per-label confidence.

Acceptance criteria:

• Sequential-fact fixture classified as supersedes with zero LLM calls
• Genuinely overlapping-interval contradiction still escalates to LLM correctly
• PII detection: >95% recall on structured PII, NER-based detection for names/addresses
• Redaction happens strictly before embedding (pipeline ordering test)

---

Track D — Read/Ranking Pipeline & Hot Cache

Goal 3.D: Semantic search, composite scoring, and hot cache

Milestones:

• 3.D.1 — Semantic search read path: find_similar via VectorStore.search(), GIN full-text fallback when vector results < K.
• 3.D.2 — Composite scoring v2: Category-conditional recency decay. Formula weights unchanged (0.40/0.25/0.20/0.10/0.05), only recency component is category-conditional. Regression test proves old factual outranks fresh episodic at equal relevance.
• 3.D.3 — Hot cache (Redis sorted set): ZADD/ZREVRANGE per agent, top-50 capacity, TTL-based eviction. Cache score uses category-conditional recency for consistency.

Acceptance criteria:

• find_similar uses HNSW index (confirmed via EXPLAIN ANALYZE)
• Cross-org result leakage impossible (integration test)
• Hot cache score formula reuses CATEGORY_HALF_LIFE_DAYS

---

Track E — Exit Gate

Goal 3.E: Prove the composed system enforces isolation and meets quality bars

Milestones:

• 3.E.1 — Memory security integration suite (TestMemory_ISO_*): 10+ cross-tenant isolation cases against real Postgres+pgvector+Redis.
• 3.E.2 — End-to-end memory lifecycle test: Full write/dedup/conflict/retrieve round trip via HTTP API. make e2e-smoke-p3-memory.
• 3.E.3 — Retrieval-quality benchmark sign-off: Recall@10, p50/p95/p99 latency, precision@k. CI regression gate.
• 3.E.4 — ADR consolidation: Phase 3 memory ADRs published and indexed (next free numbers after ADR-0047; 0040–0046 are Phase 2.5).
• 3.E.5 — Phase 3 exit audit: Severity-tiered gap register (P0–P3), same format as Phase 1 exit audit.

Acceptance criteria:

• All ISO-* cases pass with real containers
• memory-security-integration CI job required on main
• Benchmark baseline committed; CI fails on regression
• Zero open P0 gaps in exit audit