A Neuro-Cognitive Agent

What is this?

Most AI agents are stateless — each conversation starts from scratch. The Neuro-Cognitive Agent (NCA) models how the human brain actually works: parallel subconscious processing, emotionally-weighted memory, and a conscious mind that reasons over all of it.

It perceives before it thinks

Every message is parsed for intent, emotional intensity, and entities — before any reasoning begins. Like the thalamus filtering sensory input, NCA’s perception layer decides what matters.

It has a subconscious

Six specialized “subminds” run in parallel on Haiku 4.5 — metacognition, deductive reasoning, emotional coherence, graph context, anticipatory retrieval, and neocortical consolidation. A basal ganglia gate decides which ones fire.

It remembers like you do

Memories aren’t just stored — they decay over time, get boosted when recalled, and are retrieved through six different brain wave modes depending on context. A knowledge graph links entities across conversations.

It builds a mental model of you

Over time, individual memories consolidate into narrative arcs, which merge into a person schema — a persistent mental model of who you are, what you care about, and how you think. Just like a close friend would.

How It Thinks

Every turn follows an 8-step cognitive loop — modeled after the flow of information through the human brain. Here’s what happens when you say something.

1

Perception

Haiku parses the message for intent, emotional intensity (NRC lexicon), entities, and urgency. Decides the retrieval strategy.

“I’m worried about the Portugal trip” → intent: emotional_sharing, intensity: 0.65, entities: [Portugal trip]

2

Drain prior cues

Collect subconscious cues from the previous cycle — metacognition confidence, emotional reads, deductive insights. These inform the gate.

Metacognition: “complexity: medium, confidence: 0.7”

3

Push state → Subconscious

The basal ganglia gate evaluates perception signals and selects which subminds fire. A greeting activates only metacognition. A complex synthesis question activates all six.

Emotional input → gate fires: metacognition + emotional_coherence + graph_context

4

Retrieve memories

Brain wave mode selected by time gap and intent. Vector search finds relevant memories, then CA3 spreading activation pulls in connected memories through entity links. NTAS scores decay over time.

Beta mode → 7 memories; “Elena” entity → 2-hop spread finds Portugal trip + wedding

5

Assemble context block

Memories, subconscious cues, entity facts, person schema, and pacing guidance are assembled into a working memory context block with proportional token budgets. The density dial adjusts based on schema presence.

8 memories (60% budget) + 3 cues (25%) + schema (15%) + pacing: “reflective, ~250 words”

6

Cerebellar check

A forward model (Haiku) predicts whether the assembled context is good enough. If confidence is below threshold, it triggers re-retrieval or memory trimming before Opus fires.

Context confidence 0.4 < 0.6 → re-retrieve with broader query → confidence 0.78

7

Conscious reasoning

Opus 4.6 receives the full context block and produces structured cognition: thoughts, reasoning, plans, self-criticism, emotional read, and the response shown to the user.

Opus weighs career anxiety memories + schema “values autonomy” → warm, specific career advice

8

Encode & learn

The exchange is encoded as a new memory with temporal, emotional, and entity bindings. Climbing fiber signals calibrate the gate for next time. Prediction errors accumulate toward schema updates.

New memory stored; gate error: emotional submind over-predicted → tonic adjusted −0.05

Brain Map

Every numbered marker maps a real brain mechanism to its NCA implementation. Hover over a marker for details.

Medial Sagittal Section — numbers reference table rows below

Prefrontal
Cortex Thalamus Basal
Ganglia Amygdala Hippocampus Entorhinal VTA Angular
Gyrus Cerebellum Brainstem Cortical Oscillations (distributed)

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

n

Implemented

n

Planned

Brain illustration: Patrick J. Lynch & C. Carl Jaffe, Yale School of Medicine — CC BY 2.5

#	Brain Region / Mechanism	How It Affects Humans	Neuroscience Function	NCA Layer	When It Fires	Example	Phase
Sensory Gating & Perception
1	Thalamus	Filters sensory overload so you focus on what matters	Sensory gating & filtering	Perception	Every turn — first step of cognitive loop	“Tell me about Portugal” → intent: question, entities: [Portugal], urgency: 0.7	1
2	Thalamic reticular nucleus	Ignore background noise while reading	Attentional gating of working memory	Conscious	Context block exceeds token budget	12 memories + 5 cues exceed budget → lowest-scoring memories pruned; cues protected	2
Emotional Processing
3	Amygdala	You flinch before you think — gut-level alarm	Emotional salience detection	Perception	Every turn — NRC lexicon scan	“I’m terrified about the diagnosis” → intensity 0.82 (fear + anticipation dominant)	3
4	Amygdala → Spotlight	In crisis you recall vivid emotions, forget peripherals	Emotional hijack of retrieval	Memory	emotional_intensity > 0.7	Intense grief → Spotlight mode: 5 emotionally-similar memories, no spreading activation	3
5	Amygdala → Anticipatory suppression	When flooded emotionally you stop planning ahead	Emotional focus inhibits prediction	Subconscious	retrieval_hint = “spotlight”	Emotional crisis → anticipatory submind gets −0.5 suppression, stops pre-fetching	3
6	NRC Word-Emotion Lexicon	“Home” feels different than “domicile”	Plutchik 8 emotions + 2 sentiments	Subconscious	Emotional coherence submind fires (intensity > 0.3)	“I’m so proud of Lily” → joy 0.7, trust 0.5 → two-channel blend with Haiku analysis	3
Memory Encoding & Retrieval
7	Hippocampus (DG/CA3)	A smell triggers a childhood memory	Pattern completion retrieval	Memory	Every turn — after perception	“What was that restaurant?” → vector search finds Cantinho do Avillez from 80 memories	1
8	CA3 recurrent collaterals	Friend → their partner → the restaurant you all visited	Spreading activation (2-hop)	Memory	Entities found in initial retrieval (capped at 20)	“Elena” in memory → hop 1: “Portugal trip” → hop 2: “Maria’s wedding”	2
9	Hippocampal unified encoding	You remember what, when, where, and how it felt	Bind temporal + contextual + experiential + associative	Memory	After every agent response	Career anxiety talk → stored with timestamp, entities [Marcus, Dev], valence −0.3	1
10	Hippocampal replay	Sleep replays the day, strengthening what matters	Theme extraction via replay to neocortex	Subconscious	Consolidation submind fires (depth ≥ 3, info load accumulates)	5 turns about career → “Theme: tension between IC identity and leadership aspiration”	3
11	NTAS temporal decay	Forget last Tuesday’s lunch, remember your wedding	Salience decay + reconsolidation boost	Memory	During retrieval scoring	3-month-old memory (salience 0.6) decays to 0.41; re-accessed memory gets 1.15× boost	1
12	Cortical oscillations	Alert focus retrieves differently than drowsy free-association	Brain wave state-dependent retrieval	Memory	Brain wave mode selected by time gap + intent	Quick follow-up → Gamma (7 memories); “trace the arc” → Delta (12 memories, wide spread)	1–3
13	Entorhinal cortex	You know a phone number without reliving the day you learned it	Entity property binding	Memory	Perceived entities have graph nodes — fires during retrieval	“How’s Elena?” → graph returns: lives in Lisbon, works at CERN, married to Miguel	2
14	Entorhinal → Context	Hearing “Elena” instantly surfaces her key attributes	Entity facts surfacing	Conscious	Entity facts exist in graph — injected into context block	Elena mentioned → context block gets “Elena: physicist, Lisbon, partner: Miguel” as working memory	2
Cognitive Control & Gating
15	Basal ganglia (GPi/SNr)	Brain doesn’t analyze grammar when someone says “hello”	Tonic inhibition of subminds	Subconscious	Every turn — gate evaluates all subminds before firing	“Hey Koda” → only metacognition fires; deductive, graph, anticipatory all inhibited	4
16	Striatal MSNs	Several clues accumulate into a sudden realization	Multi-input integration → graded disinhibition	Subconscious	Multiple signals combine — entities + complexity + emotion accumulate activation	“Compare Elena’s career to mine” → entity (0.3) + question (0.25) + complexity (0.2) = 0.75 > threshold	4
17	VTA → D1 dopamine	Curiosity and surprise make you more mentally flexible	Arousal lowers firing threshold	Subconscious	Emotional intensity > 0.5 or novel entities detected	Surprise question about forgotten topic → arousal 0.7 lowers thresholds by 0.15, more subminds activate	4
18	VTA → mPFC dopamine	“Connect the dots” activates narrative-building circuits	Circuit-specific consolidation facilitation	Subconscious	Synthesis request or high-complexity with person schema present	“How has my thinking on career changed?” → consolidation submind gets +0.3 boost, fires earlier	5
Consolidation & Schema Formation
19	Systems consolidation	Over weeks, isolated experiences become “my year abroad”	Hippocampal → neocortical cascade	Memory	Session ends or narrative arc threshold reached	8-turn career chat → session summary → arc: “Growing tension between IC work and leadership” → schema update	5
20	mPFC schema formation	You carry a mental model of close friends	Person mental model (McClelland 1995)	Memory	Enough narrative arcs accumulate (≥3) or bootstrap from rich initial history	After 5 sessions → person schema: “values autonomy, anxious about leadership, close to Elena and Lily”	5
21	mPFC → Context block	Your mental model of someone shapes every interaction	Schema injection into working memory	Conscious	Person schema exists — injected into every context block	Schema “values deep work, anxious about promotion” appears in Opus context → shapes tone and advice	5
22	vmPFC schema-driven cognition	You reason from your model of a friend, not by replaying every talk	Reason from mental model, not from scratch	Conscious	Schema present + many memories retrieved → density dial shifts to schema-lean	Schema covers career context → density: “lean on schema, prioritize recent episodes” → fewer memories needed	5
Executive & Higher-Order Cognition
23	Prefrontal cortex (dlPFC)	“Wait, that doesn’t follow” — you catch logical gaps	Logical structure analysis	Subconscious	Gate activation > threshold for deductive submind (question or high complexity)	“Why did the project fail?” → deductive submind cue: “causal chain: timeline pressure + scope creep + no escalation”	1
24	Prefrontal metacognition	You know when you’re confused — thinking about thinking	Reasoning quality monitoring	Subconscious	Always — bypasses basal ganglia gate entirely	Low-confidence question → metacognition cue: “complexity: high, confidence: 0.4, suggest clarification”	1
25	Prefrontal cortex (executive)	The inner voice that deliberates and forms a considered response	Deliberate conscious reasoning	Conscious	Every turn — after all subconscious cues and memories assembled	Opus receives context block with 8 memories + 3 cues + schema → produces structured cognition with speech	1
26	Working memory capacity	Juggle ~4 things; more than that and you drop details	Cowan’s ~4-item limit	Conscious	Context block assembly — proportional token budgets enforced	150k token budget: memories get 60%, cues 25%, schema 15% → overflow items pruned by relevance	3
27	Predictive coding	Walking into a kitchen you pre-activate “stove, fridge, sink”	Proactive pre-fetch of associations	Subconscious	Conversation depth ≥ 2, not in Spotlight mode	3rd turn about Portugal trip → anticipatory cue pre-fetches: “likely topics: food, Elena, Lisbon weather”	2
28	Medial temporal lobe network	You navigate who knows whom, who’s connected to what	Entity relationship traversal	Subconscious	Entities detected in perception (entity count > 0)	“Tell me about Elena and Miguel” → graph traversal: Elena—PARTNER→Miguel, Elena—WORKS_AT→CERN	1
29	Angular gyrus semantic priming	“How’s the restaurant search?” — you know they mean Belcanto	Implicit entity detection	Memory	No explicit entities in input, but recent context contains relevant ones	“Any updates?” after discussing Lisbon → resolves implicit entities: [Lisbon, Belcanto, Elena]	5.2
Homeostatic Regulation
30	Adenosine sleep pressure	The longer you’re awake and learning, the more you need to sleep	Information-load consolidation trigger	Subconscious	Cumulative info load exceeds threshold during conversation	10 dense turns → info_load 8.5 → consolidation submind gets +0.4 boost, triggers mid-session summary	6a
Cerebellar Prediction & Calibration
31	Cerebellar forward model	You know the cup is heavier than expected before you think about it	Haiku evaluates context quality before Opus fires; re-retrieves or trims memories if confidence < 0.6	Subconscious	Before Opus fires — Haiku evaluates assembled context quality	Context has 3 memories but low relevance → confidence 0.4 < 0.6 threshold → re-retrieves with broader query	7.0
32	Climbing fiber error learning	Missing three free throws in a row changes how you shoot the fourth	Error-driven tonic calibration — gate learns which subminds matter per user	Subconscious	After each turn — compares predicted vs actual response quality	Emotional submind predicted “high need” but user wanted facts → error signal adjusts tonic level −0.05	7.1
33	Prediction error accumulation	“I keep misjudging her” triggers revising your mental model	EMA error tracking — 3+ consecutive mismatches trigger schema reconsolidation	Memory	EMA error > staleness threshold for 3+ consecutive turns	Schema says “anxious about leadership” but user is now excited → 3 mismatches → triggers schema rebuild	7.2
34	Cerebellar temporal sequencing	You know when to pause, when to be brief, when to elaborate	Response pacing & fluency prediction	Conscious	Every turn — analyzes message length ratio, emotion trajectory, topic continuity	User sends 5-word follow-up after long message → brief-warm register, ~50 words target	7.3
Planned
35	dlPFC relational processing	Comparisons and conditionals demand more cognitive effort	Deductive relational complexity modulation	Subconscious	Planned — compare/contrast or multi-conditional detected	“How does X compare to Y given Z?” → deductive submind gets extended context window	6
36	Dentate gyrus pattern separation	You distinguish similar but different memories (today’s parking spot vs yesterday’s)	Scaling benchmark validation	Memory	Planned — near-duplicate memories detected at retrieval time	Two lunch meetings at same café → separator distinguishes Feb vs March event by context	6

Perception

Conscious (Opus 4.6)

Subconscious (Haiku 4.5)

Memory (ChromaDB + Neo4j)

N Implemented

N Planned

Benchmarks

NCA is tested against a raw Opus 4.6 baseline in a 12-question blind A/B evaluation. An independent Opus judge scores both responses across six dimensions, without knowing which is NCA. Results below are averaged across 4 benchmark runs (48 total questions).

28–17

NCA vs Baseline wins (48 questions)

7–4

Latest run (best: 9–3)

36

Brain mechanisms implemented

Dimension Scores (NCA vs Baseline)

Averaged across 4 benchmark runs. Scale: 1–10. Delta shows NCA advantage.

Contextual Memory

+0.8

Relational Continuity

+0.6

Emotional Attunement

+0.4

Naturalness

+0.4

Connective Reasoning

+0.2

Practical Depth

+0.1

NCA score (bar width)

Delta vs baseline (right)

By category (across 4 runs)

Needle: NCA 9 – BL 4 – Ties 1

Specific fact recall from months-old conversations — restaurant names, employee counts, book titles. This is the architecture’s core strength: NTAS retrieval + entity binding finds what context-stuffing can’t.

Evolution: NCA 8 – BL 0

Tracking how opinions and feelings changed over time. Undefeated. The consolidation pipeline (session summaries → narrative arcs → schema) gives NCA a structural advantage no baseline can match.

Synthesis: NCA 8 – BL 3 – Ties 1

Connecting threads across conversations — infrastructure plans, blog arcs, trip logistics. CA3 spreading activation and graph traversal surface connections the baseline misses.

Emotional: NCA 2 – BL 4 – Scale: NCA 1 – BL 4

NCA’s weakest areas. On purely emotional questions, the architecture sometimes over-retrieves or over-structures. On large-scale personal growth questions, the baseline’s simpler approach can surface more detail. Active area of improvement.

Use Cases

NCA’s brain-inspired architecture creates capabilities that a stateless LLM simply cannot match. Each use case below shows which mechanisms power it, and a side-by-side scenario comparing NCA to a raw Opus baseline with the same conversation history pasted into context.

Personal AI Companion

Core Strength

A companion that genuinely remembers your life — names, dates, evolving preferences, and emotional context — across months of conversation. Not a chatbot that starts fresh every session.

Hippocampal Encoding NTAS Temporal Decay Person Schema Narrative Arcs Amygdala Gating Anti-Withholding Guard

Scenario: “What restaurants have I been excited about?”

UserWhat restaurants have I been excited about lately? I need dinner ideas.

NCAYou’ve been on a real food exploration kick. That Thai place on Valencia — Kin Khao — you went twice in January and said the northern curry was the best you’d had in years. You also mentioned Flour + Water after Sarah’s birthday in December, specifically the fresh pappardelle. And there was that new Ethiopian spot, Barcote, that you bookmarked after Marcus recommended it but haven’t tried yet. For tonight, Kin Khao if you want a sure thing, Barcote if you want the adventure.

BaselineI don’t have specific information about restaurants you’ve been excited about, since I don’t retain memories between conversations. Could you remind me of some places you’ve enjoyed? I’d be happy to help with dinner suggestions based on your preferences.

Therapeutic & Coaching Support

Emotional Intelligence

Tracks emotional patterns over time, recognizes growth, and responds with calibrated warmth. Knows when to listen, when to challenge, and when to celebrate progress — because it remembers the full arc.

Emotional Coherence Submind Spotlight Retrieval NRC Emotion Lexicon Schema Staleness Detection Narrative Arc Consolidation Response Pacing

Scenario: “I’m thinking about going for that promotion.”

UserI’m thinking about going for the director role. Part of me feels ready, part of me doesn’t.

NCAThat ambivalence makes complete sense given where you’ve been. Six months ago you were telling me you felt invisible in leadership meetings — that was right after the Q3 reorg when your scope got cut. But since then you led the platform migration, you started mentoring Priya and James, and last month you said you finally felt like “the person people come to.” The readiness is real. The doubt might be an echo of how you felt before, not where you are now.

BaselineIt’s completely normal to feel ambivalent about a big career move. Going for a director role is a significant step. I’d suggest reflecting on your strengths, recent accomplishments, and areas where you’ve grown. What specifically makes you feel ready, and what’s driving the hesitation?

Executive Assistant with Memory

Organizational Power

Tracks projects, people, decisions, and open threads across weeks. Synthesizes status from scattered conversations. Connects dots that a human assistant might miss because the context was spread across too many touchpoints.

CA3 Spreading Activation Entity Graph (Neo4j) Anticipatory Retrieval Delta Brain Wave Mode Entorhinal Entity Binding Session Summaries

Scenario: “What’s still open from the board prep?”

UserWhat’s still open from the board prep work we’ve been doing?

NCAThree things still open. First, the ARR slide — you said on Tuesday you were waiting for Chen to finalize the Q4 numbers, and you hadn’t heard back as of Thursday. Second, the competitive landscape section — you drafted the Snowflake comparison but said you wanted to add the Databricks pricing data that Julia’s team pulled. Third, you mentioned wanting to rehearse the “why now” narrative because last quarter Marcus gave you feedback that the strategic framing was too reactive. Want to run through that narrative now?

BaselineI’d be happy to help with board preparation. Could you share what topics you’ve been working on? I can help organize your materials, review slides, or prepare talking points for any outstanding items.

Research Partner

Connective Reasoning

Remembers every paper discussed, every hypothesis explored, and every dead end hit. Connects ideas across sessions and fields — surfacing the citation from three weeks ago that’s suddenly relevant to today’s experiment.

Graph Context Submind Deductive Reasoning Submind Theta Brain Wave Mode Metacognition Submind Consolidation Pipeline Basal Ganglia Gating

Scenario: “I’m seeing weird results in the attention heads.”

UserI’m seeing something weird — the middle attention heads are activating more than the early or late ones on our new dataset. Any thoughts?

NCAThat mid-layer activation bump is interesting. It reminds me of the Olsson et al. induction heads paper you flagged three weeks ago — they showed mid-layer heads doing the heavy lifting for in-context learning. Your new dataset is mostly few-shot classification examples, right? If the model is learning task structure from the examples rather than relying on pretrained patterns, you’d expect exactly this: early heads handle syntax, late heads handle output formatting, and the mid-layer heads are where the actual in-context reasoning happens. It also connects to your hypothesis from the February experiments about attention entropy — you predicted mid-layer entropy would be higher for novel tasks. Might be worth checking that metric on this dataset.

BaselineThat’s an interesting observation. Middle attention heads activating more than early or late ones could suggest several things. In transformer architectures, different layers tend to capture different types of information. Could you share more details about your model architecture and the characteristics of the new dataset?

Customer Success & Account Management

Relational Continuity

Maintains deep context on each account — stakeholders, pain points, usage patterns, past issues, and relationship dynamics. Turns every interaction into a continuation, not a cold start.

Person Schema (mPFC) Entity Binding VTA Dopaminergic Gate Boost vmPFC Density Control Climbing Fiber Error Learning Beta Brain Wave Mode

Scenario: “Acme Corp wants to talk about renewal.”

UserAcme Corp’s CTO just scheduled a renewal call for next week. What should I know going in?

NCAKey context for Acme: their CTO is Raj Patel, and he was frustrated in October when the SSO integration took three weeks longer than quoted. That got resolved and he said he was “cautiously optimistic” in the December QBR. Usage has been strong — they expanded from 50 to 180 seats after the product team adopted the platform in November. Two risks: Raj mentioned evaluating a competitor (DataStack) at the November happy hour, and their VP of Engineering, Lena Torres, has been quiet since the January reorg — you haven’t had a direct touchpoint with her. I’d lead with the seat growth as validation, address the SSO incident proactively to show you learned from it, and try to loop Lena in before the call.

BaselineTo prepare for the Acme Corp renewal call, I’d suggest reviewing their current contract terms, usage metrics, any open support tickets, and recent communications. It would also be helpful to identify their key stakeholders and any competitive threats. What information do you have about their current sentiment?

Development Phases

Each phase added new brain-inspired mechanisms and was validated by benchmark testing. The score shows NCA wins vs baseline losses in blind evaluation.

1

Phase 1 — Foundation

Benchmark: 6–4–2

Perception, hippocampal encoding & retrieval, NTAS temporal decay, 6 brain wave modes, conscious reasoning (Opus), subconscious workers (Haiku). The basic cognitive loop.

2

Phase 2 — Retrieval Enhancement

Benchmark: 6–6–0

Entorhinal entity binding, CA3 spreading activation (2-hop), token budgets, anticipatory submind. Memory retrieval became associative instead of just vector-search.

3

Phase 3 — Cognitive Refinement

Benchmark: 8–3–1

Amygdala emotional gating, Spotlight retrieval mode, NRC lexicon, neocortical consolidation submind, density dial. The first major benchmark jump.

4

Phase 4 — Cognitive Control

Benchmark: 8–4–0

Basal ganglia gate with graded inhibition, striatal MSN integration, VTA dopaminergic modulation. Not every question needs every submind.

5

Phase 5 — Narrative Consolidation

Benchmark: 8–3–1

mPFC person schema, VTA→mPFC gate boost, vmPFC density, anti-confabulation guard, angular gyrus implicit entity detection. The agent builds a mental model of you.

6

Phase 6a — Adenosine Sleep Pressure

Information-load tracking triggers consolidation when conversation density accumulates. The agent knows when it’s time to summarize and consolidate.

7

Phase 7 — Cerebellar Prediction

Benchmark: 7–4–1 (avg across runs)

Forward model context prediction, climbing fiber error learning, schema staleness detection, response pacing. The agent predicts, evaluates, and self-corrects. NCA leads in 5 of 6 dimensions.

...

Future — Phases 6b+

dlPFC relational complexity, dentate gyrus pattern separation, memory scaling validation at 5000+ memories.

Architecture

Four layers, modeled after the human brain’s information flow. ~4,500 lines of Python across 30+ files.

Perception

Thalamic input parsing

Haiku 4.5 extracts intent, emotional intensity, entities, and urgency from every message. Decides retrieval strategy before anything else fires.

Subconscious

6 parallel subminds + basal ganglia gate

Metacognition, deductive reasoning, emotional coherence, graph context, anticipatory retrieval, neocortical consolidation. Each runs on Haiku 4.5 in a separate process. The gate selectively activates based on input signals.

Memory

ChromaDB vectors + Neo4j knowledge graph

NTAS temporal scoring, 6 brain wave retrieval modes, CA3 spreading activation, entity property binding, multi-scale consolidation (session → arc → schema).

Conscious

Opus 4.6 deliberate reasoning

Receives assembled context block with memories, cues, schema, and pacing guidance. Produces structured cognition: thoughts, reasoning, self-criticism, and speech.

Tech Stack

Python 3.11+ Anthropic SDK Claude Opus 4.6 Claude Haiku 4.5 ChromaDB Neo4j FastAPI Pydantic PyYAML pytest uv (Astral) multiprocessing