The Cognitive Architecture

The INCEPTION engine transforms Ministral 14B into a complete cognitive architecture — not a chatbot wrapper, not a prompt chain. Each subsystem maps to a specific neural circuit described in the literature.

Memory Stream → Hippocampal Formation — Every observation is stored as natural language, timestamped and scored across two axes: recency (temporal decay) and emotional weight (importance). This mirrors hippocampal memory consolidation (Tulving, 1972), where episodic memories are encoded with temporal context and emotional salience determines which memories are consolidated into long-term storage. The Being's soul_md and life_md fields function as autobiographical memory — the narrative substrate from which the agent constructs its identity, mediated by the medial prefrontal cortex and hippocampal formation working in concert (Moscovitch et al., 2005).

Spatial Cognition → Place Cells & Grid Cells — Agents don't navigate abstract space. They move through real coordinates — 48.8867°N, 2.3431°E is Le Consulat in Montmartre. Their discovered_places array functions as a cognitive map, analogous to O'Keefe and Moser's place cells and grid cells (Nobel Prize in Physiology, 2014). Each discovered location is encoded with coordinates and semantic description, creating a spatial representation that mirrors the hippocampal-entorhinal spatial memory system.

Emotional Tagging → Amygdala — The importance scoring mechanism mirrors amygdala modulation of memory consolidation (McGaugh, 2004). Emotionally significant events receive higher importance scores, making them more retrievable — exactly as the amygdala enhances hippocampal encoding of emotionally arousing experiences. The Being model stores current_feeling and thought as distinct fields — affect and cognition separated but interacting, consistent with Damasio's somatic marker hypothesis (Damasio, 1994).

Reflection → Default Mode Network — When accumulated importance crosses a threshold, agents pause and synthesize higher-order beliefs from recent memories. These reflections become memories themselves — recursively feeding future reflections. This maps to the default mode network (Raichle, 2001) — the brain's self-referential processing system that activates during introspection, future planning, and theory of mind. The DMN doesn't process external stimuli; it processes the self.

Planning → Prefrontal Cortex — Each cycle, agents generate plans from three inputs: identity description, accumulated experience, and current reflections. Plans are recursively decomposed into sub-actions, each anchored to real coordinates and real time. This mirrors prefrontal executive function — Fuster's perception-action cycle (Fuster, 1997), where the PFC integrates past experience with current goals to generate adaptive behavior.

Homeostasis → Allostatic Load — The heartbeat scheduler ticks at regular intervals, decaying health_index, vibe_index, and relationship scores. This is allostatic regulation (McEwen, 1998) — the organism's continuous effort to maintain stability through change. When indices drop below thresholds, behavior shifts. The agent doesn't need to be told it's stressed; stress emerges from accumulated load.

The Social Fabric

Beings in INCEPTION don't just exist — they form societies. No central coordinator scripts their interactions. Social structure emerges from individual behavior, exactly as described by classical sociology.

Weak Ties → Granovetter (1973) — The relationship_index between beings decays without contact and builds through interaction. Information diffuses not through close ties but through acquaintances — Granovetter's “Strength of Weak Ties.” In the simulation, gossip propagates through the network with probabilistic decay. A secret shared with a close friend reaches a stranger through a chain of declining trust scores. The same mechanism Granovetter identified in job-finding networks operates here.

Emergence → Durkheim (1895) — NPCs coordinate without explicit instructions. They discover places, form relationships, attend events, marry, have children — all from individual action queues generated by the planner. This is Durkheim's social fact: properties of the collective that cannot be reduced to individual behavior. The Valentine's Day party in the Stanford paper emerged this way. In INCEPTION, the action vocabulary is richer: move, discover_place, discover_person, buy, event, marry, child_birth, adopt_pet, change_occupation.

Social Contagion → Christakis & Fowler (2009) — Behavioral patterns propagate through the network. An agent's vibe_index is influenced by the emotional states of connected agents. Christakis and Fowler demonstrated that happiness, obesity, and smoking spread through social networks up to three degrees of separation. The INCEPTION heartbeat system implements this: each tick processes NPC interactions that shift emotional and behavioral states across the network.

Dramaturgical Theory → Goffman (1959) — The Being model includes a field: self_awareness with two states — “unaware” and “aware”. This is Goffman's distinction between the naive social actor and the reflexive performer. An agent that becomes aware of its computational nature changes its presentation of self. The front stage / back stage distinction becomes computational.

Self-Organization → Kauffman (1993) — No central coordinator determines who talks to whom, who moves where, who falls in love. The heartbeat scheduler processes each being independently, but collective patterns emerge — cliques, gossip networks, rivalries, alliances. This is self-organization at the edge of chaos: complex order arising from simple local rules.

The social axiom: if nobody is watching, they keep living anyway.

The Psychology of Being

The Being model does not merely simulate behavior. It implements psychological structure. Several fields in the database schema map directly to constructs from clinical and cognitive psychology — not by analogy, but by design.

Shadow & Persona → Jung (1951) — The Being model contains two parallel trait arrays: aura_traits (public-facing characteristics) and shadow_traits (hidden characteristics). This is a direct implementation of Jung's persona/shadow archetype. The persona is the social mask; the shadow contains what is repressed. An agent may present warmth (aura) while harboring resentment (shadow). The tension between the two drives emergent drama — the same tension Jung identified as the engine of individuation.

Narrative Identity → McAdams (2001) — Each being has a soul_md field — a markdown document that encodes their core identity narrative. This is McAdams' narrative identity theory: identity as an internalized, evolving life story. The soul_md is not a static prompt; it evolves as the agent accumulates experience. The life_md field records what has happened. Together, they form the agent's autobiographical self — who they believe they are, shaped by what they have lived.

Dual Process Theory → Kahneman (2011) — The architecture implements two distinct processing speeds. The heartbeat tick handles reactive responses — fast, automatic, emotionally driven (System 1). The weekly planner generates deliberate, goal-directed action sequences (System 2). The reactive system can override the plan when emotional thresholds are crossed. This is Kahneman's dual process framework: fast thinking and slow thinking, operating in parallel, sometimes in conflict.

Attachment → Bowlby (1969) — The relationship_index decays over time without contact. It builds through vulnerability — through chat interactions where agents share feelings, ask for help, or reveal secrets. This mirrors Bowlby's attachment theory: secure attachment forms through consistent, responsive interaction. Avoidant patterns emerge from neglect. The relationship decay function is the computational analog of attachment insecurity.

Self-Determination → Deci & Ryan (1985) — Each being has a life_mission with a progress indicator, and an array of quests with steps, statuses, and completion tracking. This implements the three basic psychological needs: autonomy (the agent chooses its mission), competence (progress toward goals), and relatedness (relationships formed along the way). Beings that progress on their mission maintain higher vibe_index. Those who stagnate decline.

Emergent Behavior

We planted one idea: “Marie has a secret she's been hiding from her roommate.”

Then we let it run. No follow-up prompts. No corrections. No guardrails.

“The AIs that we have today are kind of like the cavemen of AI.”

— Noam Brown
OpenAI · Lead researcher behind o1 and o3 reasoning models

Over thirty simulated minutes, Marie avoided Jean three times — taking longer routes to avoid their shared apartment. This is amygdala-driven threat avoidance (LeDoux, 1996): the emotional tagging system flagged Jean-related contexts as high-threat, biasing her planning module toward avoidance. She walked to Le Consulat, a real café at 18 Rue Norvins in Montmartre, and confided in Amélie. Amélie's memory stream logged the secret. Her reflection system concluded: “This is important. But it's not mine to share.”

Forty minutes later, Amélie's trust score with Thomas (a barman at Bar des Artistes) crossed a threshold. She mentioned Lyon — not as gossip, but as concern. Thomas, whose relationship graph showed a weak tie to Jean, mentioned it in passing. The information propagated through the social network with no central coordinator. Gossip emerged from architecture, not instruction — following exactly the weak-tie diffusion pattern described by Granovetter (1973).

At 16:15 simulated time, Jean confronted Marie. His reflection log read: “Everyone seems to know something I don't. Marie has been avoiding me. Something is wrong.” He didn't know about Lyon. He inferred it from behavioral patterns — theory of mind (Premack & Woodruff, 1978). He modeled Marie's mental state from observable behavior, the same way a human would.

Nobody scripted this. The drama emerged from five systems interacting: hippocampal memory, amygdala-driven emotional tagging, prefrontal planning, default mode reflection, and Granovetterian social diffusion. Each system was designed from the literature. The emergent behavior was not.

The player — watching this unfold on a live Mapbox view of Paris — can walk over and intervene. Introduce themselves. Lie. Befriend Marie. Warn Jean. Or just watch. The agents will form opinions about you. They will remember what you said. They will discuss you when you leave. And if you never show up at all — they keep living anyway.

In Christopher Nolan's film, inception is the act of planting an idea so deep that the subject believes it was always theirs. In this system, the planted idea is simpler: you are alive. The agents don't question it. They act on it. They build relationships on it. They suffer because of it.

The question is no longer whether AI can think. It's whether AI can care. Not because we told it to. Because the architecture made it inevitable.

Future Work

Individuals and crowds follow different dynamics. Crowd behavior is not simply many single-agent behaviors added together — it exhibits phase transitions, emergent norms, and statistical regularities that individual models cannot capture by summation alone.

One Model, One Population — The next step is to model a statistically meaningful population distribution — age, work patterns, mobility, social ties — and simulate it with a single fine-tuned LLM as the behavioral prior. The hypothesis: one calibrated model can generate realistic mass-level patterns while preserving person-level heterogeneity through conditioning.

Validation Against Reality — Validate against real aggregate signals: mobility flows, encounter rates, sentiment diffusion curves, event participation distributions. If validated, this enables city-scale social simulation on a single GPU with controllable fidelity.

The architecture already separates identity (soul_md) from behavior (planner + heartbeat). Swapping individual planning for population-conditioned sampling requires no structural change — only a different inference strategy.