Synthetic Cortex defines how thinking is structured, not what a model is trained on. It attaches to an existing language model and organises inference through parallel, specialised subsystems that all read from one shared neurochemical state. This page shows the pipeline as far as we publish it.
Tap any node. Solid lime lines carry the main signal, cyan lines carry emotional state, grey lines are internal hand-offs. Red nodes were closed by the agent that prepared this page. Dashed nodes are branches we have not disclosed yet.
Gate definitions, the conversion from affective coordinates to neurochemical levels, the coupling between state and generation parameters, the exact channel set of the hidden-context manager, and several modules in development are not published. The diagram is accurate in shape and deliberately incomplete in detail.
WhyThe value of the architecture is in its orchestration. The components are public science; the way they are wired together is not.
Most systems change what a model sees or what it has memorised. Synthetic Cortex changes the conditions under which it thinks.
| Dimension | Prompted LLM | RAG pipeline | Fine-tuned model | Synthetic Cortex |
|---|---|---|---|---|
| Decides what the model sees | The prompt | Similarity to the query | Already in the weights | A neurochemical state, gates and associative recall |
| Reasoning starts | When asked | When asked | When asked | When an internal gate crosses its threshold |
| New knowledge lives | In the context window | In an external index | In changed weights | In layered memory, parallel spaces and hidden context (L2: latent space) |
| Base weights | Frozen | Frozen | Modified | Frozen |
| Retrieval timing | — | Once, before inference | — | Re-evaluated as inference evolves (L2) · emotion-weighted (M1) |
| Role of emotion | Tone, if any | None | Style learnt from data | Central regulator of priority, gating and consolidation |
| Forgetting | When the session ends | Rarely | Risk of catastrophic forgetting | Homeostatic, tier by tier, leaving traces |
| Explainability | Low | Cited sources | Low | Gates crossed, channels fired, per-module logs |
Prompt engineering works on what the model observes at the surface. The cortex works one level down: on how the model relates, organises and navigates meanings it has already inferred. Tokens are entry points into that space, not the objects of reasoning.
A conventional data pipeline engineers features and lets correlations emerge from volume. Implicit relation engineering shapes the relational geometry itself — which concepts resonate, which decay — so new information lands in an existing network of meaning instead of being learnt from scratch.
Agent frameworks route by rules and planners. Here the same input can fire different modules on different days, because routing depends on an evolving internal state with momentum, coupling and balance — closer to a regulated system than a flowchart.
Each phase addresses a specific limitation of current language models and adds a capability on top of the previous ones, without growing the core parameter count. Open any phase for its dossier.
Neurochemical signals modelled mathematically and mapped onto continuous control variables that reshape inference.
CompletedKnowledge embedded during interaction while the base model stays frozen.
In researchStructured knowledge injected into latent space during inference, without scaling parameters.
CompletedReasoning drawn from the model's own activations instead of imposed prompt structures.
Largely completeMeaning as trajectories on a curved surface rather than distances in a flat space.
In developmentControlled traversal of concept space beyond the immediate task.
In developmentThe same concept interpreted through different cultural lenses.
In developmentLimbic-2 is the current emotional engine. Its predecessor, Limbic v1, was retired and archived; the present design goes beyond representing affect and doubles as a communication protocol between cortices. In the main system, parallel auxiliary computations run during inference across the later transformer layers — where abstract patterns have already formed — and the contextual signal extracted there is analysed on valence, arousal and dominance. The dominant context is amplified, alternative contexts are kept as controlled deviations, and the result is re-expressed as thirteen neurochemical values.
Each cycle integrates with the previous one in a way that lets older cycles fade gradually rather than averaging them away, so the state has memory without being stuck. The output is a set of emotional loads that every other module is normalised against.
Emotional loads guide internal learning, latent injection and activation steering; they shape activation-based reasoning, default-mode association and cultural resonance. Weights produced by the model are deliberately modulated with these loads, triggering different probability protocols and shifting the overall reasoning flow.
The same values set the thresholds for memory synchronisation, arousal regulation, salience filtering, episodic updates, contextual persistence and logical consistency — and for any external module a developer attaches. The system stays in balance because every operation is computed relative to its current state.
The mapping that turns valence, arousal and dominance into thirteen channel levels, and the coupling between those levels and generation parameters, are withheld in this release.
See the public interaction matrix and the live frequency lab →
Language models typically process information in a flat, Euclidean space where distances are straight lines and activation spreads globally: when one concept lights up, distant and irrelevant concepts are pulled into the computation too. A manifold organises information along a curved surface. Points that look far apart in vector space can be close along the surface, and what matters is proximity along it.
The cortex keeps two such spaces in parallel. A state computed in one is not copied into the other as raw data; it crosses as an emotional context and a priority signal. Each point interacts only with its local neighbourhood, so the system reasons over a small relevant region instead of the whole space — which is also how the cost of manifold methods is kept down. Model weights stay fixed; only the path the data takes is curved.
A semantic core is extracted from internal activations and projected along directions that correspond to different conceptual or cultural frames, then mapped back to text. Abstractions emerge from relational convergence rather than from memorised sequences.
Same geometry, different job. Emotional loads from the current input, recent dialogue and episodic history are mapped onto every point of A. When accumulated load crosses a threshold, B can redirect A's reasoning trajectory.
The mobile cortex keeps every record in two parallel spaces. When a topic comes up, emotionally intense records close to it come forward, and the Reflection takes them into account — with a ceiling on how strongly that shows.
See it in the atlasChain-of-thought prompting asks a model to write its reasoning out. That improves local coherence but stays at the level of text, and it remains tied to the distributions of the training data. In the main system, reasoning is grounded in the model's real-time activations: the cortex observes which semantic relations the model itself activates and raises their neighbouring clusters into the active context. Retrieval is not a single step before inference; targets are re-evaluated as decision-layer activations evolve, tying vector proximity and decision dynamics into one flow.
In the emotional engine, the trigger for this mode is chemical: when the glutamate-analogous channel crosses its gate, chained reasoning engages; when the prolactin-analogous channel crosses its own, the loop is closed. No prompt tells the model to "think step by step".
No weight updates, no fine-tuning; the architecture monitors existing representations and builds relational structure on top of them. Particularly suited to research-heavy domains and to agent reflection phases that need adaptive context reconstruction.
Activation-guided selection can drift into confirmation bias if early anchors over-constrain retrieval. Monitoring adds latency, mitigated by parallelism. Quality depends on the embedding pool behind it.
Instead of storing information in weights or passing it as raw text, structured knowledge is injected into latent space during inference, where it interacts with internal representations as if it were already encoded. Core parameters remain unchanged, so cost and instability stay low.
Without access to a model's internals, the mobile cortex arranges data on the device with small embedding models and injects it as hidden context. Concepts that sit far apart in the model's meaning space are brought into one window — the indirect route to the same effect.
How M1 differsRicher relational training makes models more creative and less consistent. Most systems settle on a compromise. The cortex separates the two: an analytically oriented layer produces structured reasoning and extracts conceptual anchors; a specialised associative layer, trained on deliberately dense cross-domain relations, expands only within those anchors; the analytical layer then re-grounds the result. Expansion mode is entered only when synthetic affective signals cross their thresholds.
Because the creative layer fires on affective thresholds, enrichment appears spontaneously where context calls for it — not on every turn.
Completed. Information access restricted for commercial confidentiality.
Training dynamics push culturally contingent concepts — legitimacy, respect, authority, family, freedom — toward one dominant statistical interpretation. The result is abstraction loss: the model stores an average where a society has a frame. The cortex targets the relational geometry around a concept, reshaping how activation propagates through its neighbours so that interpretation shifts with the active cultural resonance profile. Emotional loads keep the module dormant unless a dominant perturbation calls for it.
In Generation Network, a first consumer-facing expression of this idea is the Time Machine: an idea is read through how different historical periods would receive it.
The main system connects to open models and reads weights during inference to synthesise its neurochemical state; it can then steer the model's geometric latent space directly. A phone cannot carry that processing volume, and the mobile build was designed to keep personal mental models on the device. So M1 removes a large part of L2 and re-engineers the rest.
The outcome is not L2 on a phone. It is a system that reaches several of the same effects indirectly — and does far more than an average chat layer. Every M1 module is listed in the Cortex Atlas.
An attempt to abstract, mathematically, the functional processes that give rise to reasoning in biological cognition — learning, using what was learnt, adapting, solving problems — and to unify them under one architecture that is open in two directions: special-purpose "brain regions" for specific needs, and broader cognitive capability on the same infrastructure.
The project does not aim to build artificial general intelligence or an independent intelligent system. Its neurochemistry is a set of mathematical models. Benchmarking is ongoing; claims on this site describe design intent and internal observations.