Identity needs story
The identity formation article introduced the identity vector: a slowly-updating representation of behavioral tendencies. The vector is numerically precise and useful for comparison and calibration, but it is not sufficient for reasoning about continuity. A vector of six numbers does not tell you what kind of system you are dealing with, what it has been doing, or where it is heading.
Narrative adds the missing dimension. By organizing identity-relevant experiences, goals, failures, and revisions into autobiographical structure, the narrative engine gives the system a compressed model of its own history and trajectory. This model serves as context for future reasoning and as evidence for constitutional safety checks.
What Experiment 25 revealed
Experiment 25 produced a striking before-and-after comparison of the narrative state across a simulated incident lifecycle.
After five baseline rounds (settled, routine work), the narrative described the dominant traits as stabilityScore, curiosity, and explorationPreference. The system's characteristic posture was stable, inquisitive, and willing to explore. Coherence was 0.786: the system's recent behavior matched its self-model closely.
After ten consecutive outage rounds (sustained high contradictionRisk = 0.9, low reinforcement = 0.1), the picture changed completely. The dominant traits became caution, verbosity, and toolDependence. Themes: outage, incident, critical, risk monitoring. The stabilityScore had fallen to 0.353. Coherence had dropped to 0.270.
The narrative engine also produced an explicit summary statement noting "low coherence." This is the identity engine reporting, in its own terms, that its self-model is under stress.
Autobiographical threads
The engine synthesizes identity-relevant memories into narrative threads. A thread may describe a recurring strategy, a long-running goal, a repeated failure mode, or a stable preference. Threads are not summaries of individual events; they are patterns across events.
An example thread might be: "The system has consistently applied caution-first reasoning when confronted with authentication-related risks, retrieving incident memories from the outage window before proposing actions." This thread is not stored anywhere as a raw fact; it is synthesized from the pattern of retrieved memories, action choices, and outcomes across many sessions.
Threads bind events across time. This gives later reasoning systems a way to perceive continuity rather than isolated experiences. When the arbitration system considers a proposal about an authentication risk, the thread about caution-first reasoning for authentication risks is context that can raise or lower the proposal's coherence score.
Narrative coherence and coherence drops
Narrative coherence measures how well recent behavior aligns with the established threads. The coherence drop in Experiment 25 (0.786 to 0.270) happened because the outage rounds produced behavior (maximum caution, near-zero exploration, complete focus on incident signals) that diverged from the established threads about a curious, exploratory system.
A coherence drop is not automatically a problem. The system's behavior during the outage was appropriate: high caution and incident focus are correct responses to a genuine crisis. The coherence drop is a measurement of how far the current behavior has moved from the prior self-model, which is useful information regardless of whether the movement is correct.
Low coherence becomes a signal to downstream systems. The constitutional engine (described later) monitors identity drift. A large coherence drop combined with a high drift magnitude from the baseline identity vector can trigger a quarantine: the current identity state is flagged for review before it is committed to the active self-model. This is exactly what happened in Experiment 26, where the post-outage identity's drift of 0.266 from baseline exceeded the maxIdentityDrift = 0.2 invariant.
Future-self projection
Beyond synthesizing past behavior, the narrative engine projects likely future identity states from current goals and reinforcement trends. If the current trend is toward increasing caution and decreasing exploration, the projection can estimate where the identity vector will be in ten, twenty, or fifty more sessions under the same signal distribution.
These projections help evaluate whether proposed actions support the kind of system the architecture is becoming. An action that is individually reasonable might, if repeated, push the identity toward a future state that is undesirable (extreme caution, near-zero exploration, loss of calibration on novel situations). The future-self projection makes this trajectory visible.
This longitudinal dimension extends arbitration beyond immediate utility. The question is not only whether an action works now, but whether repeated actions of that kind would preserve the behavioral character that makes the system reliable and adaptive. A system that always chooses the safest option becomes unable to handle genuinely novel situations; a system that always explores becomes unreliable for high-stakes routine work. The projection helps maintain the balance.
Narrative revision and prior state preservation
Belief updates and policy changes can revise the narrative. If the system's self-model says "I prefer cautious recommendations" but recent successful episodes involved creative risk-taking, the narrative should update to reflect the new evidence.
Revision is not deletion. The prior narrative state is preserved with a timestamp so later audit can reconstruct the history of self-model changes. This preservation serves the same purpose as retirement in the forgetting system: the audit trail is intact even when the active belief has changed.
The engine preserves prior narrative states selectively. Threads that have been stable for many sessions carry more evidence and are not overwritten by a few contradicting episodes. Threads that are recent or weakly evidenced can be revised by relatively few new data points. This asymmetry prevents short-term fluctuations from rewriting stable long-term patterns.