ADR-0013: AISME Phase 5 — Predictive Coding Narrative Self¶

1. Context¶

In cognitive psychology, the "narrative self" represents an agent's continuous, autobiographical identity over time. Under the predictive coding paradigm, this narrative identity acts as a top-down generative model that constantly predicts upcoming observations, actions, and user interactions. When incoming events deviate from predictions, precision-weighted prediction errors update the narrative self model.

2. Problem Statement¶

Autonomous agents lack a cohesive self-model that persists and adapts across interaction sessions. Without predictive self-modeling, agents exhibit behavioral incoherence, forget their own established perspectives, and fail to track shifts in their relationship with users. Spector needs a continuous narrative self engine that updates through hierarchical predictive coding.

3. Decision Drivers¶

• Autobiographical Coherence: Maintain an evolving identity state vector and narrative summary across multi-turn sessions.
• Predictive Coding Architecture: Implement hierarchical top-down prediction and bottom-up error propagation.
• Plasticity vs. Stability Dilemma: Prevent catastrophic identity drift while allowing authentic character growth.
• Off-Heap Identity Persistence: Store self-model state vectors in dedicated off-heap memory segments.

4. Considered Options¶

Option 1: Static Prompt Injection¶

• Description: Hardcode agent identity in system prompts and inject fixed persona text.
• Advantages: Simple configuration.
• Disadvantages: Static and brittle; cannot learn from interactions or adapt to evolving user relationships.

Option 2: Hierarchical Predictive Coding Self-Engine (Selected)¶

• Description: Maintain an off-heap NarrativeSelfState vector representing core beliefs, personality traits, and autobiographical milestones. At each turn, generate top-down predictions of user responses. Compute prediction error \(\epsilon = y - g(\theta)\); when precision-weighted error exceeds an epistemic threshold, update narrative traits via Kalman-filtered Bayesian updates.
• Advantages: Produces organic, authentic behavioral evolution; mathematically grounded in predictive coding; highly compact off-heap footprint (2 KB).
• Disadvantages: Requires calibrating prediction error learning rates to avoid identity oscillations.

5. Decision Outcome¶

Chosen Option: Option 2 (Hierarchical Predictive Coding Self-Engine).

Positive Consequences¶

• Consistent autobiographical memory and character voice across long-horizon interactions.
• Quantitative measurement of conversational surprise via prediction error tracking.
• Self-model updates occur out-of-band without degrading dialogue response latency.

Negative Consequences & Trade-offs¶

• Extreme conversational shocks require dampening to prevent personality instability.

6. Pros and Cons of the Options¶

7. Implementation Plan¶

1. Phase 1: Define NarrativeSelfState record and off-heap layout in spector-kernel.
2. Phase 2: Implement PredictiveSelfEngine and error computation in memory/spector-memory/aisme/self.
3. Phase 3: Add NarrativeSelfBiasRelay to modulate recall scoring using autobiographical relevance.
4. Phase 4: Implement periodic identity checkpointing in ReflectDaemon.

8. Code Reference & Verification¶

• Primary Module(s): memory/spector-memory, synapse/spector-synapse
• Key Packages: com.spectrayan.spector.memory.aisme.self, com.spectrayan.spector.synapse.persona
• Classes: PredictiveSelfEngine.java, NarrativeSelfState.java, IdentityUpdatePolicy.java
• Verification Tests: NarrativeSelfContinuityTest.java, PredictiveErrorUpdateTest.java