Synthetic Cortex is an external cognitive architecture that wraps a language model and decides how it thinks: which memories surface, which distant concepts are pulled together, when to reason step by step and when to stop. Not a bigger model. A better-organised one.
Every input crosses this map before a language model writes a single word. Tap any node. Lime lines carry the main signal, cyan lines carry emotional state. Red nodes were closed by the agent that prepared this page; dashed nodes are branches not yet disclosed.
A conventional chat model receives text and predicts text. Everything it can use in a given moment has to fit inside its context window, and what gets into that window is usually decided by recency or keyword similarity. Synthetic Cortex inserts a regulatory pipeline between the input and the model.
Every input — a message, a letter, a behavioural signal — is scored on three continuous axes: valence, arousal and dominance. No emotion labels, only coordinates.
Those coordinates drive a 13-channel synthetic neurochemical state. Channels push and suppress each other, carry momentum from earlier turns, and are pulled back toward balance by a homeostatic regulator.
The state acts as a field of thresholds. Crossing one fires a process: an associative sweep, a memory consolidation cycle, a selective-perception pass, a self-review, a deep pattern inference.
What the processes return is compressed into hidden context, the model's sampling behaviour is tuned to the current state, and the output re-enters the pipeline as a new stimulus.
The loop never really stops. Inputs shift the state; the state shapes the reply; the reply shifts the state again, and that shapes the next input. Between turns, background processes keep working: records are broken into fragments, linked to each other, consolidated or forgotten, and re-weighted by how emotionally intense they were when they arrived. What reaches the model at the end is not a longer prompt. It is a more selective one.
See the full flow, including the parts we keep closed, on the architecture diagram.
Synthetic Cortex does not feel. It is a mathematical model of what the brain does while feeling: the filtering, prioritising and gating that emotion performs on cognition. The aim is the same kind of outcome the brain reaches — fast, contextual, occasionally surprising decisions — reached through a different substrate.
Emotions narrow the options. Logic justifies the choice.
Similarity search returns the records closest to a query. Associative recall also brings in records that belong to the same broader area — close in meaning but not directly on topic — whenever there is room in the context. Spreading activation across a graph of fragments lets one memory call up others.
Expected outcomeAnswers that connect topics you never placed side by side.
The neurochemical state determines what is salient and what fades. Among similar records, an emotional recall ratio works alongside semantic similarity, so the most emotionally significant memory near a topic comes forward first — and a reactive state processes the same input more strongly than a calm one.
Expected outcomeResponses that weigh what matters in this context, not what is statistically common.
Output quality depends on how much meaningful information fits in the context window. Instead of padding the prompt, the cortex pre-selects and densifies: concepts that sit far apart in the model's geometric meaning space are brought into the same window through associative context management. Pre-computed emotional baselines mean the system rarely recalculates from scratch.
Expected outcomeFewer tokens wasted on raw text, more of them spent on relationships.
An affective process and a logical process are each capable of irrational deviations on their own. Integrated, they constrain each other: associative sweeps widen the field, homeostasis and logic narrow it again. Creative expansion fires only when thresholds are crossed — neither constant randomness nor rigid determinism.
Expected outcomeUnexpected connections that still hold together.
Knowledge inside a language model lives in dense "token ecosystems" that are reachable only along certain paths. Ordinary prompting keeps hitting the statistically dominant ones. By changing the path — what enters the context, and from which associative and emotional angle — the cortex reaches regions of the model's own knowledge that normal prompting leaves dormant.
Expected outcomeRare patterns in the training data become reachable again.
Information moves through layered memory — short-term, transitional, long-term — is consolidated when it proves useful or emotionally strong, and is forgotten under a homeostatic budget, leaving a trace behind. The architecture trades instant recall for depth of integration.
Expected outcomeThe longer it runs, the further its outputs move away from a raw model's.
Each module can be developed, replaced or switched off on its own, yet every module listens to the same neurochemical state. That shared state is what turns a collection of components into one system.
coupled hormone and neurotransmitter channels
spiking neural networks — three core, seven hybrid
stacked layers across Equalia and the deep layers
manifolds — one for meaning, one for emotion
short-term, transitional, long-term — each with forgetting
channels feeding the model what it was never explicitly told
Three builds are complete. They share the same idea and differ in how deeply they can reach into a model.
The first and simplest layer. Synthesises a neurochemical state and uses its weights to trigger processes — tool calling, hidden-context injection and module activation.
The main system and the most comprehensive build. Attaches to open models and works inside inference: latent-space manipulation and activation steering, driven by weights read from the model itself.
The mobile cortex. Designated L1.5, it is in practice the leanest build in the family: many features were removed and others re-engineered so the whole cortex runs locally on a phone. Small on-device models stand in for what L2 does inside a model.
L1, L2 and M1 are the first three milestones, not the finish line. Research tracks for internal learning, manifold geometry, default-mode dynamics and cultural resonance continue inside Gen 1. Gen 2 is planned.
The first product built on Synthetic Cortex. Generation Network runs the M1 mobile cortex on the phone itself: your Reflection — a digital copy built from your letters, chats and in-app behaviour — synthesises its own neurochemical state, consolidates memories while the phone is idle, and learns to resemble you over weeks rather than minutes.
Personal mental models stay on the device by default. That constraint shaped the whole M1 build: everything had to be light enough to run locally.
The site is layered the way the system is. Start anywhere.
The full pipeline diagram, the seven research phases, Manifold A and B, implicit relation engineering, activation-as-reasoning.
ModulesEvery module in the mobile cortex, mapped to the brain process it imitates and the outcome it is built to produce.
InteractiveThe 13-channel interaction matrix and a live emotion-frequency lab you can push out of balance.
TimelineFrom Limbic v1 to Gen 1 today, and what comes next.
Plain languageSynthetic Cortex explained without jargon.
WritingEssays and updates, published on the Generation Network blog.