Zombita
Research Implementation · Production Multi-Agent Platform · Ritwik Sharma
Research Context
Zombita is the production system underlying the PhD research proposal on emergent narrative in multi-agent systems. It was not built to demonstrate the research — it was built independently, and the research question emerged from observing what it already did. The three research pillars map directly onto architectural decisions made before any formal research framing existed.
Research Pillar Mapping — Implementation to Proposal
Pillar I
Epistemic State Modeling & Information Reliability
RQ1 & RQ2: How does information reliability variance affect emergent social network structure?
Implementation in Zombita
4-layer memory architecture where each layer has distinct reliability and TTL characteristics: channel buffers (30-message RAM deque), user buffers (cross-channel awareness), conversation history (48hr TTL), and persistent reputation store (PostgreSQL). Each agent accesses different information states — partial information instantiated as a structural property, not a parameter.
Pillar II
Trust Modeling & Social Structure Propagation
RQ2: What trust mechanisms are necessary for socially coherent emergent behavior?
Implementation in Zombita
Nightly reputation pipeline reads 7 days of behavioral event logs per active user and generates a personality archetype and written AI opinion stored in PostgreSQL. Agent behavior modifies its reputation state, which modifies how the system interacts with it in subsequent cycles. Social structure emerges from behavioral history, not authored profiles.
Pillar III
Consequence Permanence & Narrative Weight
RQ3: How does irreversibility interact with information reliability to affect narrative quality?
Implementation in Zombita
Three distinct permanence mechanisms in production: 40% of every shop transaction permanently burned; debt tracked per player, never resets; session-scoped economic data never deleted. The NPC campaign framework tracks player choices with UUID-based identity persistence — choices alter available dialogue irreversibly. Consequence permanence as a felt constraint, not a design statement.
System Scale — Production Evidence
4
Memory layers, distinct reliability
4
Concurrent AI pipelines, dual-model
9
Custom Lua mods, production deployed
30+
PostgreSQL tables, designed from scratch
165
Pages of technical documentation
Key Architectural Components
Economy Engine — Emergent Pricing Pillars I · III
Closed-loop financial simulation where prices emerge from four compounding factors without admin intervention. Demonstrates emergent economic behavior driven entirely by agent interaction patterns — the same mechanism the PhD framework applies to social narrative.
Price = Base × Treasury × Demand × Wealth × Recession
40% permanent burn · 5-state treasury machine · self-correcting loop
40% permanent burn · 5-state treasury machine · self-correcting loop
NPC Campaign Framework Pillars I · II
Named agents with UUID-based identity persistence across system restarts. Scripted multi-line dialogue with persistent player choice tracking written to PostgreSQL. Group membership gating. The working agent architecture the PhD proposal formalises and extends.
9 named NPC agents · persistent UUID identity · choice history in DB · group-gated interaction
Reputation Pipeline — Behavioral Archetypes Pillar II
Nightly batch process reads 7 days of event logs per active user and generates a personality archetype and written AI opinion stored per player in PostgreSQL. Production implementation of trust propagation from behavioral signal — without direct self-reporting or authored profiles.
7-day event window · GPT-4o-mini batch · archetype + written opinion · surfaced on web profile
AI Architecture — Dual-Model Design Pillar I
Four concurrent pipelines using two models: GPT-4o-mini for high-value low-frequency interactions; Qwen3 8B local inference for high-frequency passive perception. Per-model memory scope differs — instantiating the partial information architecture at the AI layer itself.
Cloud: full personality context, 15-msg history, 48hr TTL · Local: silent failure by design