Quantum-RAG ↔ conduir-evolution — Deep Analysis (Pass 2)

Quantum-RAG ↔ conduir-evolution — Deep Analysis (Pass 2)

Date: 2026-06-30
Method: 4-agent parallel review (full quantum capability catalog; conduir adaptation/eval/temporal
subsystems; conduir 72-gap memory register; conduir contracts + grafosa + plugin seams).
Relationship to Pass 1: This corrects and deepens QUANTUM_x_CONDUIR_MEMORY_ROADMAP.md.
Read this second.

0. What Pass 2 changed (the headline corrections)

Pass 1 over-credited the quantum side on adaptation and under-counted both sides' surface area.
Deep delve findings:

1. conduir is already mature where Pass 1 saw opportunity. It ships (default-on or flagged):
contextual Thompson-sampling bandits, online-gradient RRF autotune, z-score drift detection,
a write-time VoI scorer, bi-temporal validity windows (point-in-time graph queries), and a
rank-aware eval harness with Mem0/Letta adapters (QAEval + external_benchmarks).
→ *The quantum adaptive_selector, drift_compensation (Kalman), bandit/A-B, and VoI modules are
NOT differentiators. conduir is equal-or-ahead. Drop them from the value story.
2. The quantum repo is much bigger than "7 methods" — ~40 capabilities. Several were invisible in
Pass 1 and do fill real conduir gaps (below).
3. Most of conduir's 72 open memory gaps are trust/safety/lifecycle (Phase 0), not retrieval-algorithm
gaps. Quantum-inspired methods land in Phase 3–4 (retrieval quality + scale), strictly after*
the safety/eval floor. Positioning matters: this is not an immediate-win story.
4. There is a clean, low-risk integration seam nobody flagged before: the shadow embedding lane
(care_shadow_embedding.py) — read-only, off-path, scope-partitioned, output-tagged shadow=True.
It is purpose-built for exactly this kind of A/B experimentation.
5. Embeddings are NOT first-class in any canonical contract (implementation-private to the dense
index). Combined with gap AO-9 (no portable embeddings / vendor lock-in), this makes
model-agnostic HDC/VSA hypervectors strategically interesting, not just a retrieval trick.

1. Corrected capability ledger (supply vs. demand)

1a. Quantum side — the ~40 capabilities, bucketed by conduir relevance

Genuinely additive to conduir (conduir lacks or is weak):

NOT differentiators (conduir already has, equal/better): adaptive method selection, Kalman drift,
contextual bandit, A/B framework, write-time VoI, deterministic entity extraction (conduir has regex NER),
HyDE (conduir has it, flag-off), bi-temporal supersession model.

Not relevant to conduir (no QPU; keep in sandbox): native Grover, Quantum-KNN, quantum-kernel
attention on hardware, QRAM, ML quantum error-mitigation, drift/pulse/circuit control.

Concept worth learning (explains the whole bridge): Sparse Distributed Memory / Modern Hopfield
Networks ≡ Transformer attention (ADVANCE_RETRIEVAL §3) — the math linking associative memory (HAM/HDC)
to attention. This is why the quantum-inspired memory work and conduir's retrieval converge.

1b. conduir side — demand register (72 gaps), condensed, with quantum-supply match

Critical reframing: Phase 0 blockers (TS-1..9, LC-1/2/6, AO-3/4) are governance/plumbing, not
algorithms. Quantum-inspired methods cannot help there and must wait behind them.

2. The integration seams (now concrete)

The deep pass pinned down exactly where new algorithms attach:

1. RAG plugin Protocols (symposium/rag/plugins.py): Chunker / Embedder / Retriever / Reranker —
single-port, manifest-driven (component.json + handler.py). A quantum-inspired retriever/reranker
is a drop-in here.
2. MemoryTier protocol + MemoryRouter (symposium/memory/plugins.py): 5 tier kinds; new backing
store implements write/read/query/evict.
3. ★ Shadow embedding lane (care_shadow_embedding.py) — the recommended first home.
Read-only, off hot-path, scope-partitioned, output-tagged shadow=True, already used for a deferred
GNN verdict. Drop a quantum-inspired retriever/embedder here, score in parallel, compare on the eval
harness — zero production risk, no version-bump/re-encode needed.
4. RRF ensemble as a scoring leg — add a new channel's scores into the fusion pool (the same way
graph/HyDE channels attach), behind a flag, A/B vs baseline.

Constraints any new algorithm must honor (non-negotiable):

• Tenancy/scope is enforced at tier/session level (tenant::agent::user), not in the plugin —

the algorithm must filter internally (ADR-CON-3995 scope model).

• Poison/quarantine carried in metadata; algorithm must respect it.
• Embeddings stay implementation-private; canonical contracts (memory_recall_result_v0,

OutcomeEvent, MemoryRecord) carry no vectors — so a new embedder is decoupled, but a new
vector store (HDC/holographic) means re-implementing the _DenseIndex abstraction (GraphKernel is an
opaque binary wheel — not pluggable at kernel level).

• No concurrent A/B of the live FAISS channel without version-bump + re-encode → that's exactly why

the shadow lane is the right first step.

3. Did we miss something? — yes, five things (the payoff)

1. conduir's maturity inverts part of the Pass-1 story — adaptation/eval/temporal are done; stop
pitching quantum bandit/drift/VoI.
2. Multimodal memory is a clean white-space. conduir is text-only; the quantum repo already has a
working multimodal ingestion schema (ASR/audio/image → unified events). Highest-novelty direction.
3. Hypergraph (n-ary) relations — conduir's graph cannot express "A and B jointly caused C"; the
quantum repo can. Relevant to RQ-8 multi-hop.
4. The portability angle (AO-9 + embeddings-not-first-class). HDC/VSA hypervectors are model-agnostic
and composable → a path to vendor-neutral embeddings conduir explicitly wants but hasn't designed.
5. DC-3 paraphrase dedup is an open, explicitly "non-blocker / research-grade" conduir gap — i.e. a
sanctioned, low-stakes place to learn-by-building with HDC/VSA against a measured problem
(precision@10 ≈ 0.04 duplicate-dilution).

4. True research / learning opportunities (gaps neither side has solved)

These are the highest-learning items — open problems for both, where building = contributing:

• Generative retrieval (DSI/IncDSI) — index-as-model-weights; neither system has it. Incremental

document absorption would suit conduir's streaming memory.

• Paraphrase-robust semantic dedup at memory-write time — conduir DC-3 open; HDC bundling is a

candidate but unproven for this; no head-to-head exists.

• Multimodal long-term agent memory — almost no prior art ties ASR/vision events into a governed,

tiered, bi-temporal memory; conduir's governance + quantum's schema together is novel.

• Holographic/HDC graph store as a bounded-memory GraFOSa backend — fixes GP-2/3 (reject-on-full,

no persistence) with O(1)-ish associative graph queries; research-grade.

• Complex/phase embeddings vs MiniLM on a real memory benchmark (LoCoMo/LongMemEval) — the literature

gap the quantum research-expansion already flagged; conduir's eval harness makes it runnable → the most
publishable single experiment available.

5. Revised roadmap positioning

Quantum-inspired work is Phase 3–4, gated behind conduir's own Phase 0–2 (safety + measurement
floor). Sequenced:

• Gate (conduir-owned, prerequisite): close Phase 0 trust/lifecycle blockers; land EB-4/EB-1

(wire real LoCoMo/LongMemEval, non-empty dataset+prompt hashes, pass^k). Until this, every recall
claim is on synthetic data — do not optimize.

• Phase 3 — shadow-lane experiments (zero prod risk):
• HDC/VSA semantic-dedup scorer vs DC-3 duplicate-dilution.
• Complex/phase Embedder in the shadow lane vs MiniLM on the real benchmark (the publishable one).
• HD-NSW binary index prototype vs FAISS-HNSW on SL-1/SL-5/SL-6.
• Phase 4 — promote winners behind flags: MI-fusion replacing circular autotune (AL-1/AL-3);

hypergraph channel for multi-hop (RQ-8); edge_denoiser → conflict reconsolidation (DC-1) + poison
(TS-6); HD-graph backend for GraFOSa (GP-2/3).

• Frontier track (research/learning): generative retrieval (DSI/IncDSI); multimodal memory.

Decision gates: (a) no real benchmark ⇒ halt all Tier-A/B claims; (b) shadow-lane candidate must beat
baseline on the eval harness before any live-channel promotion; (c) complex embeddings must beat MiniLM on
LoCoMo or be archived.

6. Honest caveats

• "Quantum-inspired classical," not "quantum." For conduir there is no QPU benefit; framing it as

quantum repeats the claim-honesty problem conduir's own audits flag.

• Pass-2 corrects Pass-1 on conduir maturity — treat the earlier roadmap's adaptation claims as

superseded by §0/§1 here.

• Quantum modules are research-grade — need conduir's tenancy, poison-quarantine, scope, and

fail-open contracts wrapped before touching real memory.

• Two memory implementations (cortex-memory vs symposium) — target cortex-memory (canonical).
• Synthetic-benchmark trap is real and shared — both sides flag the circular-optimizer risk; the

shared discipline (BEIR/ablation + pass^k) is itself a deliverable.

• Backing literature is agent-generated — spot-check citations before betting a milestone.

7. Pointers

• Pass-1 roadmap (superseded on adaptation): QUANTUM_x_CONDUIR_MEMORY_ROADMAP.md
• Quantum capability catalog source: this review (40+ caps) — see QUANTUM_BASELINE_2026-06-30.md + quantum_rag_demo/docs/research/expansion_2026-06-30/
• conduir gap register: delivery/CORTEX-MEMORY-KNOWN-ISSUES-2026-06-18.md, …-FINAL-PRODUCT-GAP-ROADMAP-2026-06-18.md, services/cortex-memory/MEMORY-LIFECYCLE-ANALYSIS-2026-06-10.md
• Integration seams: services/cortex-memory/src/cortex_memory/care_shadow_embedding.py (shadow lane), services/symposium/src/symposium/rag/plugins.py, …/memory/plugins.py, docs/adr/ADR-CON-3995-scope-model.md, ADR-CON-3991-retrieval-dense-cascade.md
• conduir adaptation/eval: conduir_core/{contextual_bandit,drift_detector}.py, cortex_memory/{voi,memory_optimizer,external_benchmarks}.py, cortex_memory/qa_eval/

</content>