What identity means in a cognitive system
An adaptive system changes over time. Policy weights drift under reinforcement. Memory priorities shift as some experiences prove more useful than others. Reasoning strategies evolve as reflection identifies patterns in what works and what does not. Without any representation of continuity, these changes are just a sequence of local updates. The system has no way to know whether it is becoming more capable, drifting into a less reliable mode, or contradicting its own prior commitments.
Identity is the structure that tracks continuity across change. It is not a fixed persona or a constant set of rules. It is a slowly-updating model of behavioral tendencies: what the system tends to value, avoid, repeat, and revise. Identity gives the system a reference point from which to understand its own evolution.
The identity vector
The identity vector is a multi-dimensional representation of stable behavioral tendencies. Dimensions include properties like caution, explorationPreference, verbosity, toolDependence, and stabilityScore. The vector changes slowly, updated by reinforcement history and policy drift rather than by individual events.
The slow update rate is a design choice, not a technical limitation. Identity should be resistant to short-term fluctuations. A single bad outcome should not rewrite the system's self-model any more than one bad day should rewrite a person's character. The vector encodes what is persistently true about behavioral patterns, not what happened most recently.
Experiment 25 demonstrated how dramatically an identity vector can shift under sustained adverse conditions. After five baseline rounds, the vector reflected a settled state: moderate caution (0.501), good explorationPreference (0.504), stable stabilityScore (0.649). After ten consecutive outage rounds (with high contradictionRisk and low reinforcement signals on each), caution had risen to 0.957, explorationPreference had collapsed to 0.224, and stabilityScore had fallen to 0.353. The coherence measure (how internally consistent the identity vector is) dropped from 0.786 to 0.270.
This is the identity vector doing exactly what it should: faithfully encoding a behavioral reality. After sustained exposure to contradictory, high-risk evidence, the system had become genuinely more cautious and less exploratory. The identity vector was not wrong to show this; the distortion was in the situation, not the representation.
Narrative synthesis
A vector is not enough for introspection or audit. Humans asked to describe themselves do not give vectors; they give stories. The system synthesizes a narrative self-model alongside the vector: a structured description of long-running goals, characteristic strategies, repeated failures, and stable preferences.
The narrative gives later reflection and policy systems a human-readable account of continuity. It also creates a surface for contradiction detection. When recent behavior diverges sharply from the stated self-model, the divergence is detectable. A system whose narrative says "I prefer cautious, well-grounded recommendations" but whose recent actions have been consistently exploratory and poorly-supported has a coherence gap that deserves attention.
Experiment 25 made this explicit: after the outage phase, the narrative's dominant traits shifted from [stabilityScore, curiosity, explorationPreference] (describing a settled, curious system) to [caution, verbosity, toolDependence] (describing a stressed, conservative system). The narrative themes changed to outage, incident, critical, risk monitoring. The narrative explicitly noted "low coherence," which is the identity engine reporting that its self-model is under stress.
Coherence scoring and why coherence drops matter
Identity coherence measures how well recent behavior aligns with the established self-model. Low coherence does not automatically imply error. Genuine learning can change identity: a system that has acquired new capabilities or operated in a genuinely new domain may legitimately diverge from its prior self-model.
What low coherence does indicate is that the divergence is large enough to matter. A system whose coherence score drops sharply is a system that should be examined. Is the drift improving the system's capabilities? Is it drifting toward unsafe or unreliable behavior? Is the narrative model simply outdated and in need of revision?
Coherence scoring is the bridge between adaptation and stability. Without it, drift is invisible. With it, operators and the constitutional stability system (discussed later) can detect when adaptation is crossing from healthy learning into problematic instability.
Identity as input to downstream systems
The identity vector and narrative self-model are not passive records. They are inputs to later cognitive systems.
The constitutional stability layer uses the identity vector to check whether drift has exceeded safe bounds. A post-outage identity with stabilityScore = 0.353 and a drift of 0.266 from baseline exceeds the maxIdentityDrift = 0.2 invariant, triggering a quarantine. The system flags this identity for operator review before committing it to the active self-model.
The narrative engine uses identity as its starting point for autobiographical synthesis. To build a coherent story of what the system has been doing and what it is becoming, the narrative engine needs a stable (if slowly-changing) identity as an anchor.
The reflection system uses coherence drops as a signal to schedule introspection. A large coherence drop is evidence that something significant has changed and that a reflective review of recent traces is warranted.
The difference between identity and policy
Policy encodes behavioral weights: how much to explore, how much to trust established methods, how much to prefer risk-averse actions. Identity encodes behavioral character: what kind of system this is, what it tends to do across many contexts, what its stable preferences and aversions are.
Policy changes quickly, on the timescale of individual episodes. Identity changes slowly, on the timescale of many sessions. Policy is the mechanism of adaptation; identity is the record of what adaptation has produced.
This separation allows the system to adapt without losing self-knowledge. A system that changed identity as fast as policy would have no stable reference from which to evaluate whether adaptation is going well. A system that changed policy as slowly as identity would adapt too slowly to be useful. The two operate on different timescales for good reason.
The next article covers multi-agent decomposition: why a single reasoning call is insufficient for complex cognition, and how separating planning, execution, critique, memory selection, and prediction into specialized roles makes failures localizable and outputs richer.