The Store That Serialises a Sealed Hierarchy
An agent’s narrative — episodes, group experiences, themes — lives in a sealed type hierarchy. Three subtypes, each with different fields: emotional valence on episodes, consensus levels on group events, drive modulation weights on themes. Storing this in a flat key-value CBR feature map required a decision.
The existing stores (UserProfile, MentalModel, Strategy) all serialise flat records. Each field maps to a FeatureValue. NarrativeState is different: a List<NarrativeFragment> where each element could be an IndividualEpisode, a GroupEpisode, or a DerivedTheme. The CBR feature model doesn’t support nested objects.
We went with hand-rolled JSON in a single string feature, using type discriminators — "type":"episode", "type":"group_episode", "type":"theme" — and pattern-matched dispatch on deserialization. The depth-based brace-counting parser from MentalModelSchema extended naturally to handle three polymorphic subtypes with nested collections (sets, maps, enum-keyed maps). Not elegant, but consistent with the codebase and dependency-free.
The narrative feedback loop was the satisfying wiring. DriveComposer.compose() already accepted a @Nullable Map<DriveAxis, Double> for narrative modulation — added speculatively in the drive architecture. All that was needed: inject NarrativeOrchestrator into DriveOrchestrator via Instance<>, read the current narrative, call NarrativeModulation.compute(), and pass the result instead of null. One parameter that was always intended to be filled.
The norm detector was a different kind of problem. Social norms emerge from repeated multi-agent interaction — “verify before escalating” becomes a norm not because anyone programmed it, but because agents consistently follow it. The SocialNormDetector reads NormObservation cases from CBR, groups by behavioural pattern, and classifies strength: EMERGING when the pattern appears but adherence is inconsistent, ESTABLISHED when it crosses a threshold, DECLINING when a previously-established norm starts to fade. That last transition — DECLINING — requires memory of the previous state, which is why the detector caches per-tenant and compares across ticks.
Three layers of the social cognition stack now connect: narrative identity shapes drive intensity, which shapes goal proposals. An agent whose self-story includes “I help people through crises” will have amplified affiliation and competence drives in crisis contexts. The architecture was designed for this — the modulation parameter existed before the narrative system did. Filling it in was a single method reference: NarrativeModulation::compute.