NMCA: A Neurosymbolic Multimodal Cognitive Architecture

A modular, control-oriented framework for visual-symbolic intelligence
128-module specification  |  42-module cognitive core  |  May 2026 final edition

Independent architectural claim. Judge it by the tests below.

Interactive demo  ·  NMCA Blueprint PDF  ·  NMCA Addendum PDF  · 

Original 46-page blueprint (20 April 2025) — DOI 10.5281/zenodo.21972901
No patent application was filed. A copy was sent to the USPTO by email.

The public record is the Zenodo deposit.

Zenodo - 46-page first version

What is AGI?

NMCA is a documented cognitive architecture centered on a persistent, re-enterable internal spatial scene as the working state of cognition.

What NMCA Is Designed to Do

NMCA is a multimodal cognitive system centered on an Internal 3D Scene Builder as the core substrate of thought.

The architecture is designed to maintain and operate on a persistent internal spatial world rather than treating visual information as something attached to a primarily linguistic or predictive process.

Key elements specified in the original April 20, 2025 disclosure include:

• Persistent internal visual/3D scene representation used for reasoning
• Memory encoding that stores visual scenes for future recall
• Internal visual simulation synchronized with observation
• Symbolic scene representations tied to memory and belief structures
• Contradiction detection and belief updating
• Planning inside the 3D spatial model
• Simulation of routes and actions before execution
• Feedback from real observation that triggers replanning and contradiction resolution
• Multiple simulation branches
• An explicit computational loop: Input → Internal Simulation → Action → Reflective Update

A concrete example from the original disclosure is an agent that loads a location, such as a refrigerator, from prior visual memory; internally simulates the route and interaction; executes the plan; discovers that reality differs from the simulation; and generates a new subgoal.

The central architectural idea is the combination of a persistent, updatable internal 3D/visual scene with memory, belief updating, mental simulation, planning, and action. The scene is not merely an output or visualization. It is the working state on which these processes operate.

This architecture was first publicly formalized and released on April 20, 2025, in the original 46-page NMCA blueprint.

Later technical work describing related capabilities—such as persistent 3D scene memory, multi-hypothesis imagination of unobserved regions, sequential belief updating from observations, and mental simulation for embodied planning—should cite that prior public disclosure where applicable.

Final version. No further additions will be made. The author stepped away from AI/AGI research in September 2026.

The claim

Current systems are fluent predictors. They can describe a scene and generate an image of a scene. They have not been shown to keep a private, re-enterable spatial scene that remains after the original description is gone, and to use that scene as the medium of thought.

Retrieving where an object was is not the same as returning to the place where the object is.

Humans do this routinely. A short phrase is often enough to construct an internal scene that can then be inspected, moved around in, and used. Einstein’s thought experiments and Tesla’s reported mental rehearsal of machines are familiar cases of the same capacity used as a working method.

NMCA treats that capacity as architectural, not as an emergent side-effect of more tokens.

This page is a summary. The specification is the 471-page NMCA blueprint. The WebGL demo is an illustration of the distinction, not a running implementation of the 128 modules.

How it is supposed to work

Thought is not a token stream with a picture attached. The system constructs an internal scene and keeps it. Objects in that scene have identity and location.

Symbols are attached to those objects. When the system is asked where the object is, it queries the scene, not a stored sentence about the object. Reflection, planning, and memory update operate on that same state. Language is an interface into the scene, not the place thought happens.

A human reading fiction is the same distinction. The page is text. What the human reader keeps is an inner world that updates as the sentences arrive. A system that only continues the text is not doing what the reader is doing.

The first design used an explicit rendered 3D scene. The final blueprint does not keep a continuously rendered world.

The working state is an implicit scene: a structured spatial representation in code. It can be queried, updated, and re-entered without drawing it. A render is optional and on demand — for inspection, debugging, or a human-facing view.

That is a speed and resource choice. Humans do not photorealistically render an inner room either. They keep a usable spatial model and bring parts into focus when needed. NMCA is specified the same way: implicit by default, rendered only when something has to be shown.

Why 42, then 128

• Modules 1–42 — cognitive core. Visual simulation is the substrate.
• Modules 43–119 — integration with perception, latent models, learning, embodiment, and collaboration.
• Modules 120–128 — thought throttling, drift control, oversight, alignment, containment.

The later modules do not replace that core. They extend it into areas that current systems treat as training side-effects: latent-space bridges, continual learning, causal and counterfactual reasoning, embodiment after the sensor stream changes, multi-agent coordination, robustness, and long-term stability.

Length comes from granularity. Each module is specified so a researcher can inspect the function, not infer it from a trained weight matrix. A complete running system is not claimed.

The tests

Persistent scene test

Start from this prompt:

Construct a scene with a blue vase and a red apple on a wooden table.

Remove the original description. Wait. Then require the system to re-enter the scene, report where the apple is relative to the vase, change viewpoint, move an object relative to itself, take another entity’s perspective and return, and keep object identity with no new sensory input.

A system that only retrieves the stored sentence fails. The object has to still be in a place.

Embodiment

Embodiment in this architecture starts inside the scene, not in the first robot trial.

The system is given an avatar in the internal simulation and is trained there: location, reach, hiding, return, and self-dialogue against the same scene. External cameras are perception of the world. The internal scene is the mental workspace — including the system’s own position in that workspace.

After the avatar loop is stable, the same scene operations are transferred to real-world robotics tests. Modules 43–119 are the integration layer for that transfer: perception, latent models, learning, embodiment, and collaboration.

Two cameras, two jobs. Outside: what is in front of the sensors. Inside: the place thought is happening, and the self talking to that place.

The designed default is an internal scene plus external perception. If visual construction is unavailable, the same loop should still run on a non-visual spatial or symbolic workspace. That is a backup mode, not a different architecture.

Scientist Test

Give the system an unsolved problem, not a finished analogy and not a famous discovery to try to solve again independently.

Historical cases show the method, they are not the exam: Einstein’s chasing-light scene, Tesla’s assembled machine, Newton’s mountain cannon, Darwin’s tree. Those representations made the discoveries possible. If a past solved problem is given, it will only be answered from stored training text.

The scored problem must be novel and difficult enough that search, calculation, and verbal reasoning are not sufficient by themselves. The system may try those routes first. If they fail, it must construct a new visual or spatial working model and work from that.

It is scored by whether it solves the problem.

NMCA treats both tests as requirements of general intelligence: inhabit an internal world, and use it to form new representations.

How current approaches handle this

Pure scaling + scaffolding

Strength: highest current capability.

Limit: prediction, retrieval, and tools. No native requirement for a persistent scene as the medium of thought. Identity is reconstructed from records.

World models / JEPA-style predictors

Strength: reject language-only sufficiency; build internal dynamics for planning.

Limit: latent and predictive. Re-entry, identity continuity, and scene-pegged memory are not first-class by default.

Neurosymbolic hybrids

Strength: more inspectable structure.

Limit: the predictive engine usually remains primary. Symbols help; they are not the place thought happens in.

NMCA

Visual / implicit-3D simulation is the substrate. Symbols are pegged to scene entities. Identity, reflection, contradiction handling, and re-entry sit in the same loop.

On this specific goal, NMCA is the most explicit written architecture. That is a design claim. It is not a claim of empirical superiority.

Predicted walls

Each funded trajectory can keep improving capability metrics and still fail the same probe: after context is cleared, does the system know where the object is because a scene continues, or only because a sentence was stored?

• Scaling + agents: better tools and memory lookup. Presence stays reconstructed.
• World models: better “what happens next.” Still a predictor, not a place that can be re-entered.
• Memory banks: the agent queries records about a scene. It does not inhabit one.
• Hybrids: structure remains a helper unless the scene becomes the substrate.
• Robotics / SLAM: positions are known while sensors are live. The map is not turned into the place where thought continues after the stream ends.

The apple distinction

A language model can complete patterns around the word “apple,” describe fruit, or generate a picture. That is not the same as keeping an apple in a private scene and later knowing where it is because the place remained.

When the original sentence is gone, current systems typically do one of three things:

• hold the sentence in context
• store the sentence and retrieve it
• update a short-lived latent state

Fluency conceals the difference. It does not close it.

Already Done?

A robot map of a house, a next-frame predictor, or a picture of an apple is not the tests above.

Those systems track the world in front of sensors, or they store the sentence. They do not keep a private place after the words are gone.

In the public machine-architecture literature examined for this record, no document dated before April 20, 2025 specifies those tests. 10.5281/zenodo.21972901

Theoretical research blueprint for academic discussion, sandbox simulation, and human-guided exploration. Not intended for autonomous deployment.