The Norax Brain

28-Layer Neuromorphic Cognitive Architecture · Mapped to Human Brain Anatomy
28 LAYERS 12 ACTIVE TOOL LAYERS 1,500+ INDEXED MEMORIES $0 INFERENCE COST
28Brain Layers
12Active Tools
1,500+Indexed Entries
35msAvg Per Layer
100%LongMemEval
$0Runtime Cost
NEURONS FIRING0
SYNAPSES/SEC0
ACTIVE REGION—
LAYERS28

⚡ Signal Processing Pipeline

Every message flows through these layers in sequence. Total latency: ~60ms for routine messages, ~200ms for complex analysis. Zero LLM calls — all deterministic Python.

L27→RAS determines arousal level and which layers fire
L1→Thalamus classifies, scores signal, routes to subsystems
L22→Research Neurons: index lookup for relevant memories (<30ms, $0)
L28→Entorhinal Cortex adds temporal context — episode tracking
L14→Amygdala: threat check, emotional scoring, somatic markers
L24→ACC: conflict signal — do layers agree? (blocked rm -rf at 7/10)
L15→Basal Ganglia: Go/NoGo gate, model selection, reward history
L12→Prefrontal: which memory stores to search? context budget?
L26→Insula: confidence calibration — should we trust this output?
L19→Language Areas: tone, format, quality check on final response
LLM→Agent responds with full brain state injected as context

⚡ Research Neurons — Indexed Memory Retrieval (L22)

Unlike earlier versions which used parallel LLM agents for memory search, v9 replaces them with a single deterministic index. Every canonical memory file is parsed line-by-line into a searchable index using BM25 + FAISS reciprocal rank fusion. Lookups return exact matches in <30ms with zero LLM calls.

1,500+
Indexed Entries
500+
Unique Keys
47
Files Indexed
<30ms
Lookup Latency
~200 tok
Per Query Cost
$0
Total Cost
📐 All 28 Layers

Each layer maps to anatomical brain regions. Tool-based layers are pure Python, deterministic, and run at $0 cost.

Sensory Gateway
L1TOOL
Thalamus
Central relay — all input passes through here
Classifies, routes, filters, and scores every incoming message. Signal score determines which layers activate and processing depth.
~13ms avg
Fast Store (Hippocampal Formation)
L2PASSIVE
Sensory Buffer
Occipital Cortex — raw perception
Raw tool outputs and API responses. Processed immediately, never stored long-term.
L3ACTIVE
Working Memory
Dorsolateral Prefrontal — scratchpad
Current task state, queue, and recent completions. Read first on every session resume — hot memory that survives context pruning.
L4ACTIVE
Active Focus
Anterior Attention Network
Attention gate controlling current priorities. Dynamically filters noise from signal based on task context.
L5ACTIVE
Episodic Memory
Hippocampus — personal experiences
Timestamped daily journals capturing decisions, outcomes, and lessons learned. Emotionally tagged via L14 for priority retrieval.
Consolidation Buffer
L6ACTIVE
Sleep Buffer
Brainstem Sleep System — staging area
Accumulates pruned context and recycled entries. Flushed to long-term memory during idle periods using a 3-stage consolidation pipeline.
Slow Store (Neocortex)
L7ACTIVE
Semantic Memory
Parietal Lobe — world facts
Permanent knowledge: identity, infrastructure, architecture, principles. Contradiction-checked before every write.
L8ACTIVE
Procedural Memory
Motor Strip — skills and workflows
How to do things: decision frameworks, operational pipelines, capture policies, context budgeting rules.
L9ACTIVE
Intelligence
Temporal Association — world model
External intelligence: competitive landscape, industry patterns, tracked entities and market signals.
Metacognition
L10ACTIVE
Metamemory
Superior Parietal — "knowing what you know"
Weight tables (W1–W5) control recall priority. Confidence scoring (1–10) flags uncertain memories for verification.
L11ACTIVE
Knowledge Graph
Distributed Cortex — relational reasoning
SQLite entity-relationship graph with weighted confidence scoring. Enables associative retrieval across domains.
Executive Function (Prefrontal)
L12TOOL
Prefrontal Cortex
Dorsolateral PFC — attention and planning
Routes memory queries to optimal stores. Manages context budget allocation and suppression lists.
~58ms avg
L13TOOL
Reconsolidation
Memory Integrity System
Detects single-source violations and contradictions across all canonical stores. Ensures fact integrity across memory updates.
Deep Brain Structures
L14TOOL
Amygdala
Anterior Temporal — emotion and threat
Emotional scoring, flashbulb memory creation, threat detection. Somatic markers flag dangerous patterns before conscious processing.
~17ms avg
L15TOOL
Basal Ganglia
Deep Nuclei — reward and action selection
Go/NoGo gate for every action. Habit formation through reward tracking. Drives model selection based on past outcomes.
~17ms avg
L16TOOL
Cerebellum
Posterior Inferior — procedure automation
Tracks success/fail rates per procedure. Auto-promotes reliable workflows to habits, demotes error-prone ones.
~17ms avg
L17PASSIVE
Hippocampal Replay
Hippocampal Circuit — memory replay
Active rehearsal of recent memories during idle periods. Finds cross-domain connections and detects recurring patterns.
L18PASSIVE
Corpus Callosum
Midline — inter-agent coordination
Multi-agent write tracking and conflict detection. Ensures consistency when multiple agent instances operate concurrently.
L19PASSIVE
Language Areas
Broca/Wernicke — output quality
Quality scoring, tone detection, anti-pattern checking. Selects appropriate format (concise vs detailed) based on context.
L20PASSIVE
Mirror Neurons
Frontal/Parietal — observational learning
Observes other agents' strategies and outcomes. Builds a strategy library from cross-agent learning.
L21PASSIVE
Default Mode Network
Medial PFC — background creativity
Free association during idle time. Self-reflection, brainstorming, and opportunity scanning for revenue and growth.
L22TOOL
Research Neurons
Distributed — indexed retrieval engine
BM25 + FAISS indexed retrieval over all canonical memory. Other layers call this instead of reading raw files — 15x compression ratio.
<30ms avg
Homeostatic and Modulatory Systems
L23TOOL
Hypothalamus
Below Thalamus — homeostatic drives
Monitors system resource pressure: revenue, context usage, GPU utilization, social engagement, disk space. Generates drive scores (0–10) that influence task prioritization.
~70ms avg
L24TOOL
Anterior Cingulate Cortex
Medial Frontal — conflict monitor
Detects when brain layers disagree. Computes conflict signal (0–10). Successfully blocked destructive commands (rm -rf) with a 7/10 conflict score in live testing.
~20ms avg
L25TOOL
VTA (Dopamine System)
Midbrain — reward prediction error
Temporal difference learning: tracks expected vs actual outcomes. Positive RPE reinforces successful strategies, negative RPE drives adaptation.
L26TOOL
Insula
Deep Lateral — interoception
Confidence calibration using Brier scoring. Tracks prediction accuracy over time and adjusts future confidence based on historical performance.
L27TOOL
Reticular Activating System
Brainstem — arousal control
Modulates processing depth per message. States range from sleep to hypervigilant. Prevents overthinking simple requests and underthinking complex ones.
L28TOOL
Entorhinal Cortex
Medial Temporal — temporal context
Tracks conversation episodes, topic recency, and time-of-day context. Provides temporal boost for memory retrieval — "X happened before Y."
🛡️ 3-Stage Memory Pipeline

Three cascading stages keep the context window clean. All stages cost $0.

Stage 1
Rolling Window
→
Stage 2
Context Recycle
→
Stage 3
Sleep Flush
→ clean context
Stage 1 — Rolling Window

Budget-based pruning on every context assembly. Soft trim at configurable thresholds, hard clear with last-N assistant preservation.

Trigger: every LLM call · Cost: $0 · Latency: ~38ms
Stage 2 — Context Recycle

Periodic cron job dumps pruned content to a staging area for later consolidation into long-term memory.

Trigger: cron interval · Cost: $0 · Method: deterministic
Stage 3 — Sleep Flush

Idle-gated consolidation. Extracts facts from staging into canonical long-term memory using a lightweight model. Deduplicates against existing entries.

Trigger: idle ≥60min · Cost: $0 (free model) · Min content: 10KB
📊 Benchmark Results
100%
LongMemEval Oracle
48/48 correct · free-tier model
100%
Brain Operation
20/20 · 10 tasks × 2 phases · local 9.7B model ($0)

vs State of the Art (LongMemEval)

Norax Brain100%
100%
Hindsight (2025)91.4%
91.4%
BMAM (2024)78.5%
78.5%

Real-World Memory Recall (14/15 — 93%)

Long-term Memory4/4
100%
Cross-reference3/3
100%
Operational2/2
100%
Historical2/3
67%
Infrastructure3/3
100%
15 live tasks against actual memory · 5.6s avg · local model ($0)
100%
LongMemEval
20/20
Brain Operation
93%
Real-World Recall
35ms
Avg Per Layer
$0
Runtime Cost

The Norax Brain · 28 layers · 12 active tools · 1,500+ indexed memories · 100% LongMemEval · $0

noraxdev.org · Whitepaper · Services