ADR-0027: Soul-Conditioned & Salience-Modulated Personalized Dreaming¶

1. Context¶

DreamPathway delivers the 12-stage synaptic generative dreaming pipeline, Panama FFM off-heap dream journal, and FLAG_DREAMED source monitoring. In biological brains, sleep-dependent memory consolidation is inherently personalized: an individual's personal history, emotional baselines, cognitive boundaries, core values, and active concerns dictate what is replayed, how fluidly concepts recombine, and which insights survive prefrontal reality testing.

2. Problem Statement¶

Prior to this decision, DreamPathway operated on a global, homogeneous configuration. Exploration temperature, seed selection, Langevin diffusion, and Expected Free Energy (EFE) triage were identical for every agent and tenant. While Spector already possessed rich multi-soul and selective attention constructs (SoulContext, AgentSoul, UserSoul, OrgUnitSoul, TenantSoul, SalienceProfile), these were disconnected from the dreaming engine, resulting in generic, persona-blind dream consolidation.

3. Decision Drivers¶

• Personalized Memory Consolidation: Dream replay must resonate with the specific persona's core values, ethical guardrails, and emotional baseline.
• Hardware-Accelerated Seed Selection: Computing composite salience across large memory stores must utilize SIMD/GPU vectorized kernels.
• Biophysiological Faithfulness: Implement Hartmann boundary personality modulation and soul-guided Langevin stochastic differential equations (SDE).
• Hierarchical Governance: Enforce strict 4-tier governance (\(\text{TenantSoul} \succ \text{OrgUnitSoul} \succ \text{AgentSoul} \succ \text{UserSoul}\)) during dream triage.

4. Considered Options¶

Option 1: Global Dream Configuration with Prompt Injection¶

• Description: Keep dreaming pipeline parameters identical across all agents; prepend persona instructions during final LLM dream text generation.
• Advantages: No changes to mathematical diffusion or seed scoring kernels.
• Disadvantages: Fails to personalize seed selection, associative tunneling, or EFE triage; wastes compute dreaming irrelevant scenarios.

Option 2: Isolated Per-Persona Dream Engines¶

• Description: Instantiate separate dream pipeline runtimes and daemons for every registered soul.
• Advantages: Complete isolation between personas.
• Disadvantages: Excessive thread and memory resource footprint; unmanageable daemon proliferation under multi-tenant deployments.

Option 3: Context-Propagated Soul Conditioning in Unified DreamPathway (Selected)¶

• Description: Propagate SoulContext and SalienceProfile dynamically via DreamSignal through the existing 12-stage pipeline, modulating seed salience, boundary thickness, Langevin drift, and EFE triage per cycle.
• Advantages: High computational efficiency; zero daemon explosion; deep mathematical personalization.
• Disadvantages: Requires propagating soul context into all 12 dream relay stages.

5. Decision Outcome¶

Chosen Option: Option 3 (Context-Propagated Soul Conditioning in Unified DreamPathway).

Architectural Decisions:¶

D1: Multi-Soul and Salience Context Propagation¶

• Enhance DreamSignal to carry SoulContext primarySoul, List<SoulContext> soulContexts, and SalienceProfile salienceProfile.
• Thread active souls and salience profiles from SpectorMemoryBuilder, DefaultSpectorMemory, and SpectorMemoryFactory into DreamPathway.Builder.
• Preserve 100% backward compatibility when soul or salienceProfile is null.

D2: Soul-Conditioned & Hardware-Accelerated Seed Salience (SalientSeedRelay)¶

• Replace the static seed formula with a composite score evaluating recency, novelty, primary soul cosine alignment, and user salience profile semantic matching: $\(S(m) = w_r R + w_n N + w_{\text{soul}} \cos(\mathbf{e}_m, \mathbf{e}_{\text{soul}}) + w_{\text{sal}} \max_{k} \left[ \text{mult}_k \cos(\mathbf{e}_m, \mathbf{e}_{\text{interest}_k}) \right]\)$

• Strictly execute all vector similarity calculations via SIMD/GPU batch AcceleratorRegistry.getSimilarityKernel().cosineSimilarity(...).

• All weights and thresholds defined in SpectorPropertyConstants (zero hardcoded magic numbers).

D3: Hartmann Boundary Personality Modulation (RemReplayRelay / SceneConstructRelay)¶

• Derive boundary thickness factor \(\kappa_{\text{boundary}} \in [0.75, 1.35]\) from AgentSoul.emotionalBaseline() and personality traits.
• Scale effective REM temperature \(\mathcal{T}_{\text{eff}} = \mathcal{T}_{\text{base}} \cdot \kappa_{\text{boundary}}\) and Hoel regularizing noise \(\sigma_{\text{eff}} = \sigma_{\text{base}} \cdot \kappa_{\text{boundary}}\).

D4: Soul-Guided Langevin Diffusion SDE (LangevinDiscoveryRelay)¶

• Augment the continuous stochastic differential equation over the holographic memory tensor with a soul attractor potential \(V_{\text{soul}}(\mathbf{v}) = \frac{1}{2}\lambda_{\text{soul}} \|\mathbf{v} - \mathbf{e}_{\text{soul}}\|^2\): $\(\mathbf{v}_{t+1} = \mathbf{v}_t - \eta \left( \nabla E(\mathbf{v}_t; \mathbf{T}) + \lambda_{\text{soul}}(\mathbf{v}_t - \mathbf{e}_{\text{soul}}) \right) + \sqrt{2\eta \mathcal{T}_{\text{eff}}} \boldsymbol{\epsilon}_t\)$

• Bias interstitial discovery towards the agent's core purpose while thermal noise enables tunneling over high-energy barriers.

D5: Hierarchical Multi-Soul EFE Triage (EfeTriageRelay)¶

• Enforce strict 4-tier governance: \(\text{TenantSoul} \succ \text{OrgUnitSoul} \succ \text{AgentSoul} \succ \text{UserSoul}\).
• Ethical guardrail violations trigger immediate NOISE triage and Hebbian synaptic inhibition (\(\Delta w = -0.05\)).
• True IDENTITY triage outcome awarded when scene embedding resonates with primarySoul.identityEmbedding() (\(\cos \ge \tau_{\text{identity}}\)).

D6: Namespace & Soul-Aware Daemon (DreamDaemon)¶

• DreamDaemon extracts the active soul context and salience profile from DefaultSpectorMemory during periodic sleep consolidation cycles.

Positive Consequences¶

• Transforms dreaming from a generic global thermostat into an individualized, biologically faithful consolidation engine.
• Distinct agents and users with different souls produce distinct seed selections, exploration temperatures, and distilled insights from the exact same memory store.
• Zero hardcoded magic numbers—all weights and thresholds are externally configurable via SpectorPropertyConstants.
• Preserves full backward compatibility and strict FLAG_DREAMED source monitoring isolation.

Negative Consequences & Trade-offs¶

• Slight increase in seed scanning compute cost (mitigated by batch hardware SIMD/GPU cosine kernels).

6. Pros and Cons of the Options¶

7. Implementation Plan¶

1. Phase 1: Enhance DreamSignal with SoulContext and SalienceProfile fields.
2. Phase 2: Implement SIMD-vectorized seed scoring in SalientSeedRelay.
3. Phase 3: Add Hartmann boundary scaling to REM replay and soul attractor term to Langevin diffusion.
4. Phase 4: Implement hierarchical 4-tier EFE triage in EfeTriageRelay.
5. Phase 5: Validate with end-to-end multi-persona dream consolidation test suites.

8. Code Reference & Verification¶

• Primary Module(s): memory/spector-memory, memory/spector-kernel, nucleus/spector-config
• Key Packages: com.spectrayan.spector.memory.pathway.dream, com.spectrayan.spector.memory.pathway.dream.relay, com.spectrayan.spector.kernel.store
• Classes: DreamPathway.java, DreamSignal.java, DreamDaemon.java, DreamJournalMemory.java, RemDreamJob.java, DreamProperties.java
• Verification Tests: DreamPathwayTest.java, DreamGateAbortTest.java, DreamJournalMemoryTest.java