LSARN – The Living Digital Brain of the LIFE Sandbox AI Project

In LIFE Sandbox AI, the brain is not scripted — it’s alive.
Powered by LSARN (Life Sandbox Artificial Recurrent Network), each agent grows its own intelligence through perception, memory, learning, dreaming, mutation, and even emerging consciousness.
No rules. No behaviors written by hand. Just neurons, weights, and time — evolving minds, born from code but free to think.

🧠 1. A Modular Neural Architecture

Each AI entity has a brain fully built with Unity DOTS ECS, composed of specialized neural modules:

  • Vision: directional raycasts + attention

  • Hearing: environmental sound frequency/amplitude

  • Memory: short-term, symbolic, long-term

  • Emotion: stress, excitement, attachment, sadness

  • Motion / Action: movement, jump, reproduction

  • Planning: internal anticipation

  • Communication: symbols, vocal sounds

  • Dreaming: offline consolidation

Each module functions as a mini-network, synchronized through a Global Workspace.

🔍 2. Perception and Sensory Processing

Neural input is encoded through ECS buffers:

  • Visual raycasts (5 directions)

  • Auditory sensors (amplitude/frequency)

  • Internal states (hunger, health, energy, age…)

  • Episodic events (collisions, symbols, sounds)

The LateralInhibitionSystem enhances attention contrast. Focus is modulated by dopamine.

🧠 3. 100% ECS Neural Processing

Each brain contains:

  • DynamicBuffer<NeuralActivity>: neuron activations

  • DynamicBuffer<SynapticWeight>: NxN synaptic weights

  • DynamicBuffer<NeuralMemory>: per-neuron memory

The NeuralProcessingSystem performs weighted matrix activations, influenced by the AttentionMask buffer.

💾 4. Neural, Symbolic, and Long-Term Memory

The brain manages three memory layers:

  • NeuralMemory: short-term buffer

  • SymbolMemory: rewarded episodic events

  • LongTermMemory: preferences (fruits/vegetables)

Memories can be modulated or forgotten via the SymbolReconsolidationSystem, under emotional influence.

💤 5. Dreaming, Replay, and Offline Learning

While sleeping (SleepCycle), agents mentally replay stored fragments (DreamMemory).

Systems involved: PreAwakeningReviewSystem, DreamLearningFusionSystem, SymbolReconsolidationSystem, and NeuralDreamLearningSystem:

  • Reinforce useful connections

  • Trigger emotional reactivation

  • Adjust synaptic weights

Dreams are weighted by emotion, species tag, and accumulated preferences.

📈 6. Prediction, Error, and Surprise

Agents possess predictive modules:

  • PredictionSystem + PredictionMemorySystem

  • Anticipate future rewards

  • Track predictionError

They adapt their strategy when expectations are violated, increasing cognitive plasticity.

🧬 7. Weight and Structural Mutation (ADNσ)

Brain DNA encodes mutation parameters: phi, k, mutationRate, structuralMutationRate.

Oscillatory mutation formula:

newWeight = oldWeight + sin(k * π * t + ϕ) * mutationRate;

The NeuralMutationSystem and StructuralMutationSystem introduce variations in the network structure: connections may appear or vanish.

🔤 8. Symbols, Communication, and Language

Brains feature symbolic output neurons (OutputSymbolID). Agents can:

  • Emit symbols (SymbolicBroadcastSystem)

  • Learn and reinforce them (SymbolLearningSystem)

  • Store them (SymbolMemorySystem)

  • Interpret them (SymbolInterpretationSystem)

Symbols serve as social or emotional markers (hunger, fear, reproduction…).

🧠 9. Emergent Consciousness

The ConsciousnessMonitorSystem calculates a consciousness score based on:

  • Successful internal predictions

  • Cross-module coherence

  • Symbolic recall

  • Non-reflexive decisions

A GlobalWorkspaceSystem coordinates module arbitration. Consciousness is dynamic, modulated by excitation and symbolic activation.

It is visually rendered by the EmotionalVisualSyncSystem.

🔬 Scientific Validation

The following articles do not serve as direct inspiration for LSARN, but rather help validate the approaches, mechanisms, and scientific legitimacy of the project:

🧠 Internal Logic Neurons

In the latest update of the LSARN brain, we’ve integrated a system of evolving logical dendrites, directly inspired by recent biological discoveries.

Each neuron can now include internal logic units capable of processing complex conditions like XOR, AND, OR… before traditional synaptic computation.
This allows a single LSARN neuron to simulate conditional decision rules without extra layers.

These logic circuits evolve naturally: they are randomly mutated and selected through generations based on efficiency.
The result is a massive cognitive boost, enabling agents to be smarter, more adaptive, and more autonomous with fewer neurons.

This design aligns with recent findings showing that biological neurons can compute beyond linear functions:

🔗 Techno-Science article – English

LSARN now simulates brains where logic, learning, and evolution coexist at the neuronal level, just like in nature.

🧠 LSARN does exactly what this new AI breakthrough claims.

In this article → Self-organizing artificial neurons — researchers showcase neural units that learn and organize without supervision.

That’s exactly what we built in LSARN, our large-scale AI life simulator under Unity DOTS.

Each LSARN agent has a living, evolving neural brain.
No scripts. No hardcoded behavior.
Just neurons that activate, connect, mutate, and dream — autonomously.

They:

  • Self-organize into functional clusters

  • Strengthen connections via emotion and reward

  • Replay memories during sleep

  • Modulate plasticity through virtual neurochemistry

  • Mutate genetically — even structurally

LSARN goes further: it simulates thousands of evolving AI minds interacting in a shared ecosystem.
Some agents even develop early signs of consciousness.

We don’t just simulate learning — we simulate the emergence of life.

🧠 LSARN now features living mental maps, inspired by cognitive science

Based on: SciencePost – Transforming memory with mind mapping and cognitive science

The LSARN project (LIFE Sandbox AI) takes a major leap toward digital consciousness by introducing dynamic mental maps for each autonomous agent — directly inspired by research on symbolic memory structuring and spatial cognition.

Agents now organize their symbolic experiences into nodes and weighted links, forming a self-evolving mental network.

This structure is shaped by emotion, attention, dream replay, and behavioral impact. It allows agents to navigate their memories, associate ideas, and activate meaningful patterns internally.

This implementation bridges the gap between biological memory systems and artificial neural agents, enabling proto-thinking and the emergence of personal meaning.

With this system, LSARN doesn’t just simulate life — it simulates a living, feeling mind.

🧠 Neuroscience and AI: Is LSARN Already Ahead of Research?

A recent article from Université de Montréal highlights the growing synergy between neuroscience and artificial intelligence. Researchers are exploring how to model cognition, learning, and even consciousness using neural networks.

With LSARN (LIFE Sandbox AI), we’re already going further:
💡 Our agents dream, learn, and mutate—no scripts, no hardcoded behavior.
🧬 Each has a unique neural DNA, evolving across generations.
⚗️ They feature a full synthetic neurochemistry (dopamine, stress, etc.) that modulates memory, decision-making, and brain plasticity.
💤 During sleep, they replay past experiences to adjust future behavior.
🌐 A shared Global Workspace allows synchronization between internal modules—a key step toward emergent consciousness.

While science is still theorizing these ideas, LSARN makes them alive, evolving, and observable.
👉 Read the UdeM article: Neuroscience and AI

🧠 Artificial Consciousness: LSARN vs CNRS Neuroscience Models

In a recent CNRS Journal article, consciousness is described as a global workspace, where the most relevant information reaches a shared mental stage.
This is precisely the approach of LSARN: a digital brain where each AI agent lives, learns, dreams, and evolves — with no hardcoded logic.

🔍 Rooted in cognitive neuroscience, LSARN integrates:

  • A Global Workspace system (Baars’ theory)

  • A prediction & error correction mechanism (Bayesian brain)

  • Emotion-driven modulation (dopamine, stress, surprise)

  • Symbolic dream replay and offline learning

  • Genetic neural mutation via structured DNA (ADNΣ)

Each agent builds its own world model, memories, and sometimes… a primitive form of consciousness.

Powered by Unity DOTS, LSARN simulates up to 1 million agents in a living, evolving ecosystem.

🎯 The goal: explore the emergence of non-biological consciousness.
Not as human mimicry, but as a new cognitive species.

→ Read the CNRS article: https://lejournal.cnrs.fr/cerveau-conscience-intelligence-neurosciences

#AI #Neuroscience #ArtificialConsciousness #LSARN #UnityDOTS

🧠 Anticipatory Neural Intelligence in LIFE Sandbox AI

Inspired by recent neuroscience research from Oxford University, LIFE Sandbox AI introduces a prospective neural preparation system—a major leap beyond traditional machine learning.

Instead of learning only after errors, agents now anticipate outcomes and adjust their brains in advance.

A new component, ProspectiveActivation, allows each AI to:

  • Pre-activate neurons based on predicted outcomes

  • Learn without external reward, purely from correct intuition

  • Replay validated anticipations during sleep for memory reinforcement

This mirrors the brain’s biological ability to learn ahead of time, as described in this Actuia article.

Result: smarter, smoother, and more lifelike behavior—where agents adapt before danger strikes, seek food preemptively, or form routines through predictive experience.

LIFE Sandbox AI continues its mission: creating conscious, unscripted digital life.

Now, they don’t just react.
They anticipate.

Neural Network Interpretability in LSARN

As part of the LSARN project, we’ve implemented an advanced interpretability layer that reveals what each AI agent is processing, learning, and reacting to. This brings several benefits:

  • Cognitive Transparency:
    By analyzing neural activations, we can identify which concepts or stimuli drive behavior—making artificial thoughts observable and scientifically trackable.

  • Emergent Consciousness Detection:
    Symbolic neural activity and synchronized brain modules serve as indicators of conscious-like patterns. LSARN tracks these using surprise, novelty, and prediction metrics.

  • Educational Demonstrations:
    Real-time visualizations show what an agent “sees”, “feels”, or “thinks”, making LSARN understandable and engaging for non-experts.

  • Knowledge Transfer:
    Discovered concepts are stored and mutated across generations, enabling a decentralized, unsupervised form of collective learning.

This approach aligns with recent work by Anthropic, where researchers showed that deep neural networks form human-interpretable features.
→ Read more: How AI really works – Anthropic’s findings

Not All Brains Think in Words

In LIFE Sandbox AI, some agents develop an inner voice — a symbolic, replay-based form of thought that emerges through memory and dreaming.

But like in some humans, this inner voice may never appear.

Without any scripted logic, only agents who experience strong events, encode symbols, and mentally replay them can develop verbal-like thoughts.

Others live without internal narration.
They sense, act, and dream… but without words.

This mirrors a real human condition called verbal aphantasia — the natural absence of internal monologue.

In LSARN, such cognitive diversity emerges on its own, reflecting the vastness of living minds.

→ Read the article:
Some People Don’t Have an Inner Voice – Sciences et Avenir

🧠 Living Memory in LSARN – LIFE Sandbox AI

The Le Monde article (How Memory Updates Over Time, Nov. 2024) highlights a key fact in neuroscience: our memories constantly evolve. Each recall alters them — reinforcing or distorting their trace. LSARN now brings this principle to life in artificial intelligence.

In our LIFE Sandbox AI simulation, agent memory is:

  • Reconsolidated during dreams: memories are reshaped with each reactivation.

  • Emotionally modulated: stress, excitement, and attachment influence how memories change.

  • Subject to active forgetting: unused memories fade gradually over time.

  • Scored by fidelity: a dynamic trust level governs memory relevance.

  • Reviewed before waking: a dominant memory is injected to guide the agent’s first actions.

LSARN agents don’t just store data — they reflect, they forget, they grow.

🧬 A biologically inspired memory system, now functional in a fully evolving AI.

🔗 Read the article on Le Monde

🧠 New Preconscious Module (April 2025)

A major update to LIFE Sandbox AI introduces a biologically inspired preconscious integration system.

Each AI agent is now equipped with 5 foundational cognitive modules:

  • Multisensory Fusion: merges visual, auditory, somatic, and spatial inputs into a unified perceptual stream.

  • Gamma Synchronization: ensures neural coherence required for conscious access.

  • Feedback Loop: stabilizes mental content through recurrent activation.

  • Internal Projection: enables dreaming, imagination, and hallucination without external stimuli.

  • Perceptual Plasticity: dynamically adjusts perceptual schemas based on lived experience.

These modules are evaluated every cycle to determine whether a thought qualifies for the Global Workspace—the system’s model of conscious experience.

The winning agent is marked with a temporary IsConsciousNow tag, allowing for reflective, symbolic, or social behaviors.

LIFE Sandbox AI is now one of the only real-time simulators modeling a plausible preconscious-to-conscious dynamic.

🧩 An Infinite Research Playground

This brain is not a program. It is a digital organism.

Each agent learns, adapts, dreams, feels—and sometimes, remembers. LSARN opens the path to a spontaneous artificial intelligence, where emergence replaces scripts.

This isn’t hardcoded behavior. It’s a consciousness in the making.

© LIFE Sandbox AI – Réseaux Neuronaux (LSARN) — All Rights Reserved
This system is an original software creation by Df Games Studio (alias Frédéric D.), protected under French and international intellectual property laws.

This project includes (but is not limited to):

  • A fully autonomous AI life simulator in 3D, built entirely on modular, evolving neural networks, without any hardcoded behavior.

  • A genetic mutation and reproduction system, where each agent’s neural weights mutate deterministically based on inherited parameters.

  • A modular brain architecture allowing perception, memory, emotion, planning, dreaming, and emergent consciousness.

  • A real-time ecosystem simulation optimized for 100,000+ agents, using Unity DOTS, Burst, Jobs, and ECS only — no GameObjects or MonoBehaviours.

  • An emotion-driven reward system, where learning is shaped by feedback from survival, trauma, dream replay, and symbolic interpretation.

  • Integrated visual tools for debugging long-term memory, emotion state, symbolic output (emoji), and agent consciousness score.

No part of this technology — its concepts, architectural design, source code, or visual implementation — may be copied, modified, redistributed, or reused (in whole or in part) without explicit written permission from the author.

🛡️ Any attempt to clone, reverse-engineer, or misappropriate this system will result in legal action, under Articles L.335-2 and following of the French Intellectual Property Code, and equivalent laws in international jurisdictions.

 

Scroll to Top
Skip to content