Basal Ganglia
Most people who've heard of the basal ganglia heard it in one sentence: "they're involved in movement, and they break in Parkinson's." That's true and it's almost useless, because it makes the structure sound like a motor relay — a wire between intention and muscle. It isn't. The basal ganglia are the brain's action-selection and habit-learning machine. They are the reason you can drive home without remembering the drive, the reason a craving hijacks a person who genuinely wants to quit, the reason a skill becomes effortless, and the reason willpower runs out faster than habit does. They are, in the most literal neural sense, the actor inside the habit loop — the part that takes a cue, retrieves a learned action, and runs it whether the conscious "you" is paying attention or not.
This is also the structure where the gap between pop-science and real mechanism is widest. "Dopamine = pleasure" is wrong, and the basal ganglia are exactly where you see why. So let's pull it all the way down — molecule to behavior — and be honest about what's established, what's contested, and what's myth.
What it is — the deep anatomy
The basal ganglia are a set of interconnected subcortical nuclei — clusters of gray matter buried deep beneath the cortex, sitting on top of and around the thalamus, near the base of the forebrain (hence "basal"). They are not one organ; they are a committee. The core members:
The striatum — the input station, where almost everything enters. In primates it's split by a band of white matter (the internal capsule) into two visible lumps:
- Caudate nucleus — a long, curved, C-shaped structure that arcs alongside the lateral ventricle. Heavily wired to the prefrontal and association cortex. The caudate leans toward cognitive and goal-directed action selection — choosing what to do based on expected outcome.
- Putamen — sits lateral and lower, packed against the globus pallidus. Wired to motor and premotor cortex. The putamen leans toward motor learning and habit — the how, the executed movement, the automated routine.
Functionally, the caudate and putamen are the same tissue (together: the dorsal striatum), just connected to different cortical bosses. There's also a third piece — the ventral striatum, dominated by the nucleus accumbens — which is the reward/motivation arm and the hinge between this whole system and the limbic system. Hold onto the accumbens; it's where addiction lives.
The striatum's main cell, roughly 90-95% of its neurons in rodents (somewhat fewer in primates), is the medium spiny neuron (MSN) — a GABAergic (inhibitory) projection neuron studded with dendritic spines that receive glutamate from cortex and dopamine from the midbrain on the same spines. That convergence — excitatory cortical signal and dopaminergic teaching signal landing on one spine — is the molecular seat of reinforcement learning. Remember that geometry; the whole machine turns on it.
The globus pallidus — the output and relay, with two functionally opposite halves:
- External segment (GPe) — an internal hub of the "indirect" loop.
- Internal segment (GPi) — the main output nucleus. GPi neurons are tonically active — they fire constantly, ~60–80 Hz, even at rest. Their job is to continuously inhibit the thalamus. This is the single most important fact about basal ganglia function and we'll return to it: the resting state of the system is a brake held down.
The substantia nigra ("black substance" — it's literally dark, pigmented with neuromelanin), in the midbrain, two parts:
- Pars compacta (SNc) — the dopamine factory. These neuromelanin-rich neurons project dopamine up into the striatum (the nigrostriatal pathway). These are the cells that die in Parkinson's. When a Parkinson's brain is sectioned at autopsy, the substantia nigra has visibly lost its blackness — you can see the disease with the naked eye.
- Pars reticulata (SNr) — a second output nucleus, functionally a twin of the GPi, tonically inhibiting the thalamus and superior colliculus (eye movements especially).
The subthalamic nucleus (STN) — a small lens-shaped nucleus below the thalamus. The system's lone excitatory (glutamatergic) member among a sea of inhibition. The STN is the engine of the indirect pathway and the "hyperdirect" brake. It's also the most common target for deep brain stimulation in Parkinson's — stick an electrode here and you can quiet tremor in real time.
So the circuit, simplified: cortex → striatum (input) → globus pallidus / substantia nigra (output) → thalamus → back to cortex. A loop. Information from the cortex dives down into the basal ganglia, gets processed, and is fed back up to the cortex via the thalamus. The basal ganglia don't command movement directly — they release or restrain cortical and brainstem motor programs by adjusting the brake on the thalamus.
What it does
Strip away the names and the basal ganglia do four overlapping jobs:
- Action selection. At any instant, your cortex is proposing dozens of possible actions at once — reach, speak, look, step, scratch. They can't all run; many are mutually exclusive. The basal ganglia are the gate that picks one and suppresses the rest. Mink's influential 1996 "center-surround" model puts it cleanly: the system releases the chosen action while actively clamping down on the competitors. You experience the output (a clean, single movement); you never feel the suppression of the fifty alternatives.
- Procedural / habit learning. When an action sequence is repeated and rewarded, the basal ganglia gradually compress it from a deliberate, attention-hungry chain into a single automated "chunk" that runs on a cue. This is procedural memory — riding a bike, touch typing, the route to work. It's a fundamentally different memory system from the hippocampus's conscious, episodic memory, which is why an amnesiac who can't remember practicing a skill can still get better at it.
- Reinforcement learning. The system uses dopamine as a teaching signal to update the value of actions: do something, get a better-than-expected result, strengthen that action's circuit. This is where the basal ganglia and the entire field of computational reward learning meet — and it's the deepest, most validated part of the story.
- Vigor and motivation — the "go" amplitude. Beyond which action, the basal ganglia help set how much — the gain, the willingness to expend effort. This is why the same Parkinson's dopamine loss produces both difficulty initiating movement (slow, small steps) and the apathy/anhedonia that often comes with it. Dopamine isn't just selecting; it's setting the price you'll pay to act.
The mechanism — direct and indirect pathways, gated by dopamine
This is the heart of it. Get this and everything clinical falls out as a consequence.
Start from rest. The output nucleus (GPi/SNr) is tonically firing, holding the thalamus down. No movement. The thalamus, when released, would excite the cortex to execute a motor program — but it can't, because it's pinned. So the default state of the basal ganglia is no. Movement requires actively lifting the brake.
Two pathways do opposite things to that brake:
The DIRECT pathway — "GO." Cortex excites a population of striatal MSNs → those MSNs (inhibitory) project straight to GPi → they inhibit the inhibitor. GPi quiets down → it stops pinning the thalamus → thalamus is released → thalamus excites cortex → the action runs. Net effect: direct pathway activation = disinhibition of the thalamus = facilitate the selected movement. Two inhibitory steps in a row equals net excitation. This is the "go" signal.
The INDIRECT pathway — "NO-GO." Cortex excites a different population of striatal MSNs → these project to GPe (inhibitory) → GPe normally restrains the STN, so inhibiting GPe releases the STN → the STN (excitatory) drives GPi harder → GPi clamps the thalamus down even more → movement is suppressed. Net effect: indirect pathway activation = stronger brake = inhibit competing or unwanted movements. This is the "no-go" signal.
There's a third, faster route worth knowing: the hyperdirect pathway — cortex → STN directly, bypassing the striatum. It's a rapid global "stop everything" brake, important for cancelling an action you've already started (you reach for the wrong cup and yank back). Nambu and colleagues characterized this in the early 2000s.
Now the crucial part — dopamine is the switch that biases go vs no-go. And the trick is that dopamine does opposite things to the two pathways, because the two MSN populations express different receptors:
- Direct-pathway MSNs carry D1 receptors. Dopamine on D1 is excitatory — it makes "go" neurons more responsive. Dopamine turns up GO.
- Indirect-pathway MSNs carry D2 receptors. Dopamine on D2 is inhibitory — it quiets the "no-go" neurons. Dopamine turns down NO-GO.
So a pulse of dopamine does both edges of the same blade at once: it amplifies "go" and suppresses "no-go." Dopamine biases the entire basal ganglia toward acting. This is the whole logic of Parkinson's in one line: lose the dopamine, and you lose the "go" amplifier and the "no-go" suppressor simultaneously — the brake wins, movement starves. (D1/D2 segregation, established largely by Gerfen, Surmeier, and others through the late 1980s–2000s, is now well supported but not perfectly clean — a minority of MSNs co-express both, which matters for the debates section.)
Where does the dopamine itself get its instructions? This is the deepest cut, and it's the discovery that reorganized the field. The SNc dopamine neurons don't fire to report pleasure. They fire to report reward prediction error — the gap between what you got and what you expected.
Wolfram Schultz's recordings in monkeys, through the 1990s (the canonical papers: Schultz, Dayan & Montague, Science, 1997), showed it cleanly. Give a monkey unexpected juice → dopamine neurons burst. Now ring a bell before every juice → the burst migrates backward to the bell (the predictor), and the juice itself stops causing a burst, because it's now expected. And if the bell rings but the juice doesn't come → dopamine neurons dip below baseline at the expected time. A pause. That's a negative prediction error: "worse than expected."
This is exactly the temporal-difference (TD) error from reinforcement learning theory. Schultz, Dayan, and Montague made the link explicit: the dopamine signal is the teaching signal a reinforcement-learning algorithm would need. Burst = "that action was better than predicted, strengthen it." Dip = "worse than predicted, weaken it." And because dopamine differentially hits D1 (go) and D2 (no-go) MSNs, that error signal physically rewrites the synapses that select actions — a burst strengthens the "go" pathway for whatever you just did, biasing you to do it again. Reward learning is not metaphorically happening in this circuit; the circuit is literally the algorithm running in wetware.
That's why "dopamine = pleasure" is wrong. Dopamine is the prediction error and the wanting, not the liking. Kent Berridge's work (University of Michigan, 1990s–present) nailed the dissociation: you can knock out dopamine in rats and they still show the "liking" facial reactions to sweet taste — they still enjoy it — they just stop wanting it enough to work for it. Liking and wanting are separable systems. Dopamine drives wanting. Burn that into your model; almost every dopamine misconception dissolves once you do.
The landmark science
A real timeline, named:
- 1817 — James Parkinson, "An Essay on the Shaking Palsy." Described the disease before anyone knew the brain region. The eponym came later (Charcot, who refined the description and named it for Parkinson in the 1870s).
- 1919 — Constantin Trétiakoff showed that the lesion in Parkinson's was in the substantia nigra — loss of the pigmented cells — connecting symptom to structure.
- 1957–1958 — Arvid Carlsson demonstrated dopamine is a neurotransmitter in its own right (not just a noradrenaline precursor) and that depleting it caused Parkinson-like immobility in animals, reversible with L-DOPA. This won him the 2000 Nobel Prize and is the foundation of every dopamine therapy since.
- 1960 — Ehringer & Hornykiewicz found dopamine is severely depleted in the striatum of Parkinson's brains at autopsy. The smoking gun.
- 1961 — Birkmayer & Hornykiewicz gave L-DOPA to patients and saw dramatic, if temporary, reversal of rigidity. George Cotzias in the late 1960s worked out the high-dose oral regimen that made L-DOPA a real treatment — the most dramatic drug in neurology, immortalized in Oliver Sacks's Awakenings (encephalitis-lethargica patients, 1969).
- 1983 — the MPTP discovery. A batch of contaminated synthetic heroin gave a group of young California drug users sudden, full-blown Parkinsonism (J. William Langston). The toxin MPTP selectively killed SNc dopamine neurons. Tragic, but it handed researchers the first faithful animal model of the disease and accelerated everything.
- 1989–1990 — Albin, Young & Penney; Alexander & Crutcher; DeLong formalized the direct/indirect pathway "box-and-arrow" model of basal ganglia circuitry. This is the framework taught in every neuroscience course today. DeLong's monkey work predicted that the overactive STN in Parkinson's could be a surgical target —
- 1990s — deep brain stimulation (DBS) of the STN (Benabid and colleagues, Grenoble) turned that prediction into the most successful functional neurosurgery for movement disorders. A high-frequency electrode in a structure the size of a lentil can abolish tremor while the patient is awake on the table.
- 1993 — the Huntington's gene (HTT, the CAG-repeat expansion on chromosome 4) was identified by the Huntington's Disease Collaborative Research Group — one of the first triumphs of positional cloning, after Nancy Wexler's decades of fieldwork with the Venezuelan family cohorts at Lake Maracaibo.
- 1997 — Schultz, Dayan & Montague, reward prediction error = TD learning. The computational rosetta stone.
- 2005 onward — optogenetics (Karl Deisseroth, Stanford) let researchers turn the direct and indirect pathways on and off with light in living mice. Kravitz, Kreitzer et al. (2010, Nature) did the decisive experiment: stimulate D1/direct neurons → mice move more; stimulate D2/indirect neurons → mice freeze. A 2012 follow-up (Kravitz, Tye & Kreitzer, Nature Neuroscience) showed D1 stimulation is reinforcing and D2 stimulation is aversive. The model, confirmed by direct causal manipulation.
That last point matters: the direct/indirect framework isn't just a tidy diagram inferred from lesions. It's been switched on and off with light, and the predicted behaviors fell out. That's about as strong as causal evidence gets in neuroscience.
How it shapes behavior and experience
Now the part that touches your actual life.
Habits are basal ganglia recordings. Ann Graybiel's lab (MIT) ran the experiment everyone should know. Rats learn a T-maze: a tone, then run, then turn for chocolate. Early on, while the rat is figuring it out, striatal neurons fire throughout the whole run — the brain is fully engaged, evaluating every step. As the habit forms, the firing migrates to the bookends — a burst at the start (the cue) and a burst at the end (the reward), with the middle going quiet. The behavior gets chunked. The basal ganglia compress the sequence into a unit that fires off at the cue and runs to completion with minimal monitoring. This is the neural signature of "I drove home on autopilot." The cue (pulling out of the parking lot) triggers the chunk (the route), and consciousness checks out for the middle. This is the literal mechanism behind Charles Duhigg's "habit loop" (cue → routine → reward) — and the routine is a basal ganglia chunk.
Habit vs. goal — two systems, and the handoff. Early in learning, behavior is goal-directed: flexible, sensitive to whether the outcome is still worth it, run mostly by the prefrontal cortex and the caudate/dorsomedial striatum. With repetition, control shifts to habit: rigid, cue-triggered, insensitive to the current value of the outcome, run by the putamen/dorsolateral striatum. The clean demonstration is outcome devaluation (Adams & Dickinson, 1980s; striatal localization by Yin, Knowlton & Balleine, 2004–2006): feed a rat to satiety on a reward so it no longer wants it. A goal-directed rat stops working for it. A habit-trained rat keeps pressing the lever — it's running the chunk, no longer consulting whether it actually wants the thing. That's not a rat quirk. That's you finishing the bag of chips you stopped enjoying three handfuls ago. The habit system doesn't ask if you still want it. It already has its instructions.
Skill = procedural compression. Every expert skill — the surgeon's hands, the pianist's runs, the athlete's read — is the basal ganglia (with the cerebellum) having burned a deliberate, effortful, cortex-heavy sequence down to an automatic procedure. This is why deliberate practice works and why it must be effortful: you're forcing the system to encode, and the encoding only happens while the action is still under evaluative (dopaminergic) control. Once it's a habit, you stop improving — you just run it. This is the neural reason "practice doesn't make perfect, perfect practice makes perfect": automate sloppy form and you've recorded sloppy form into a structure that no longer asks for corrections.
Cravings. A cue you've associated with reward fires the ventral striatum / nucleus accumbens and triggers a dopamine surge — anticipatory wanting, not pleasure. The smell of a cigarette, the notification ping, the sight of the bar. The pull you feel toward it is the prediction-error machine screaming "reward predicted here." This is why willpower is the wrong frame: you're not fighting weakness, you're fighting a trained circuit that's faster than deliberation and was built by your own past repetitions.
When it misfires
The basal ganglia are a catalog of how the brain breaks, and each disease is a different point in the circuit failing.
Parkinson's disease — too much brake. SNc dopamine neurons die (roughly half or more of them, and around 70–80% of striatal dopamine, gone before motor symptoms appear — the system compensates until it can't). Lose dopamine → lose the "go" amplifier (D1) and the "no-go" suppressor (D2) → indirect pathway dominates → GPi over-fires → thalamus over-inhibited → movement starves. The cardinal signs (bradykinesia — slowness; rigidity; resting tremor; postural instability) are the brake winning. L-DOPA replaces the missing dopamine; it's dramatic early but the therapeutic window narrows over years (motor fluctuations, dyskinesias). DBS of the STN disrupts the runaway indirect-pathway output (exactly how is still debated). Note what this teaches: the deficit isn't "the motor cortex forgot how to move" — the programs are intact, they're just gated shut. Remove the gate (drug or electrode) and the movement is right there.
Huntington's disease — too little brake (and then collapse). Autosomal dominant, the CAG-repeat expansion in HTT. The striatum degenerates, and crucially it hits the indirect-pathway (D2) MSNs first. Lose the "no-go" suppressor → unwanted movements escape → chorea: involuntary, dance-like, writhing movements. It's the mirror image of Parkinson's: Parkinson's is too much suppression (poverty of movement), early Huntington's is too little (excess movement). As the disease progresses and the whole striatum and cortex degenerate, it broadens into cognitive decline, psychiatric symptoms, and eventually rigidity and death. Huntington's is also the cruelest natural experiment in genetics: fully penetrant at 40 or more repeats, the repeat length predicts onset age, and a child of an affected parent has a 50% chance — testable before any symptom. The ethics of that test reshaped genetic counseling.
Hemiballismus — one node, dramatic failure. A stroke in the subthalamic nucleus removes its excitatory drive to GPi → GPi under-fires → thalamus disinhibited → violent, large-amplitude flinging movements of the opposite-side limbs. A single lesion in a lentil-sized nucleus, and an arm throws itself across the room. It's a textbook demonstration that the STN's job really is to brake.
OCD and Tourette's — corticostriatal loops stuck. Both implicate basal ganglia loops failing to gate properly — in OCD, a "something is wrong" signal that won't shut off and an action (checking, washing) that won't release; in Tourette's, tics that are partially suppressible movements escaping the gate. DBS and the same circuit logic apply.
Addiction — the reward system, hijacked. Drugs of abuse spike dopamine in the nucleus accumbens far beyond any natural reward, and — critically — they don't habituate the way natural rewards do. The prediction-error system keeps screaming "better than expected" because the drug pharmacologically forces the signal. Over time, control migrates from the ventral/goal-directed striatum to the dorsal/habit striatum (Everitt & Robbins' model) — the behavior becomes compulsive habit, cue-driven, divorced from actual liking. This is the molecular tragedy of addiction: the wanting system gets trained to a stimulus it can never satisfy, while the liking often fades. People in deep addiction frequently report they don't even enjoy it anymore. Berridge's wanting/liking split, lived out.
Pharmacology, broadly: antipsychotics block D2 receptors (quieting psychosis) — but chronic D2 blockade in the striatum produces drug-induced Parkinsonism and tardive dyskinesia. Stimulants (cocaine, amphetamine) flood striatal dopamine. The entire psychopharmacology of motivation and movement runs through this circuit.
The debates and the frontier — and where pop-culture distorts it
Where pop-culture gets it wrong, flatly:
- "Dopamine is the pleasure chemical." No. It's the prediction-error / wanting signal. The entire "dopamine detox" / "dopamine fasting" trend is built on this error. You can't "deplete" or "reset" dopamine by avoiding stimulation; the term is borrowed metaphorically. The behavioral practice of reducing compulsive cue-reward loops can be useful — but the neuroscience branding is wrong, and the people selling "dopamine resets" are selling a folk theory.
- "You get a dopamine hit from your phone." Closer, but lazy. The phone is a cue that triggers anticipatory wanting via trained prediction. The variable-ratio reward schedule (sometimes there's a like, sometimes nothing) is the kind of uncertain schedule that keeps the dopamine signal firing (Fiorillo, Tobler & Schultz, 2003, found sustained dopamine activity peaks when reward probability is 50%) — same as a slot machine. That part is real and worth taking seriously. But "hit" implies pleasure; it's wanting.
- "Habits take 21 days." Myth, traceable to a misread of a 1960 plastic-surgery anecdote (Maxwell Maltz). The real number, from Phillippa Lally's 2009 study, had a median of 66 days with enormous individual range (18 to 254). The basal ganglia don't run on a calendar; they run on repetitions in a stable context with reward.
Where the real science is genuinely contested:
- The box-and-arrow model is a useful lie. The clean direct/indirect dichotomy taught everywhere is increasingly known to be a simplification. The pathways aren't fully segregated; some MSNs co-express D1 and D2; the GPe is now understood to be far more than a relay (it has distinct cell types — prototypic and arkypallidal — with their own functions, work largely from the 2010s onward). Cyriel Pennartz, Joshua Berke, and others argue dopamine's role is richer than a scalar error signal — it may also encode motivational vigor, uncertainty, and even movement initiation on fast timescales that the slow-learning-signal story doesn't capture. The 1990 model is a scaffold, not the building.
- *Does dopamine cause movement or enable it? A live argument. Some recent work shows dopamine signals ramping up during* movement and approach, not just at reward — suggesting it's doing real-time vigor/invigoration, not only after-the-fact teaching.
- How separable are habit and goal-directed systems, really? The two-system story (Daw, Dayan: "model-free" habit vs "model-based" goal-directed) is influential and well-supported behaviorally, but the brain almost certainly blends them continuously rather than flipping a switch. The arbitration mechanism — what decides which system is in charge at a given moment — is an open frontier.
- The action-selection consensus itself. Even the foundational claim — that the basal ganglia's core computation is selecting among competing actions — has critics who'd frame it more as gain control, vigor, or reinforcement than "selection" per se. The Redgrave/Gurney/Prescott action-selection model (1999) is the dominant framing, not a closed case.
The honest summary: the direction of the model — input striatum, dopamine-gated go/no-go, output that brakes the thalamus, learning via prediction error — is rock solid and causally confirmed. The details (clean segregation, dopamine-as-pure-scalar-error, hard system boundaries) are actively being rewritten. Anyone who tells you it's all settled is selling the textbook, not the literature.
How to work with it
You can't will your basal ganglia. You can train them. Everything actionable follows from one fact: this system learns from repetition-in-context plus prediction error, and once trained it runs on cues, faster than deliberation.
- Engineer the cue, not the willpower. Habits are cue-triggered chunks. The leverage is upstream — change what cues are present in your environment, and you change which chunks fire. Hide the phone, lay out the gym clothes, remove the snacks from sight. "Environment beats willpower" is literally true here: willpower is the slow prefrontal/goal system fighting the fast cued-habit system, and the fast system usually wins. So don't fight it — change its triggers.
- Repetition in a stable context is the encoder. New habits form by doing the same action in the same context, with reward, until control migrates from goal to habit. Same time, same place, same trigger. Variety is the enemy of automation. (This cuts both ways — variety is the friend of breaking automation: change the context and the old chunk loses its cue.)
- Front-load the reward to recruit dopamine. The system learns from prediction error, so a habit you want to build encodes faster if the action is reliably followed by something better-than-expected, especially early. Make the new behavior immediately rewarding (not just eventually). That's not soft — it's how you get the teaching signal to fire on the action you're trying to install.
- To break a habit, attack the cue or the prediction. You can't delete a learned chunk — it stays recorded (which is why relapse is so easy after years clean: the cue still triggers the trace). What you can do is remove the cue, or repeatedly run the cue without the reward so the prediction error goes negative and the link weakens (extinction). This is slow and the trace never fully erases — be honest about that.
- Practice perfectly, because you're recording. Whatever form you automate is the form you keep. Slow down, get it right, then let it become a habit. Automating error is worse than not automating, because now it runs without supervision.
- Respect that some of this is just biology. If the motivation/vigor feels broken — apathy, inability to initiate, anhedonia — that can be a dopamine-system state (depression, Parkinson's, burnout), not a character flaw. The basal ganglia's "willingness to pay the effort cost" is a real, sometimes-failing physical parameter. Naming it as biology is the start of treating it as biology.
How It Connects
This maps most directly to the Law of Cause and Effect — and not loosely. The basal ganglia are the brain's literal cause-and-effect engine: they take an effect (a better-or-worse-than-expected outcome, encoded as a dopamine prediction error) and use it to rewire which cause (action) you'll select next time. Reinforcement learning is cause and effect compiled into synapse. The Hermetic claim that "every effect has a cause, and chance is just a name for a law not recognized" has a striking neural correlate here: your habits feel like "just who you are," but they are nothing but the accumulated effects of past causes — repetitions you ran, rewards you got — recorded into a structure that now produces them automatically, beneath awareness. The "chance" of your daily behavior is the unrecognized law of everything you've previously reinforced. Master cause and effect at the level of the basal ganglia and you stop being the effect of your old reinforcements and start being the cause of your future ones — by controlling cues, repetitions, and rewards on purpose.
It also threads into Cause and Effect's companions: Rhythm (the system runs on neural oscillations and on the rhythm of repetition that builds a habit) · Perpetual Transmutation (skill and habit are deliberate effort transmuted into automatic procedure — energy changing form, never lost) · and the Law of Attraction in its legitimate, mechanistic reading (the trained cue-reward circuit makes you want and approach what you've reinforced — your reality bends toward your past repetitions).
Related structures and concepts: the nucleus accumbens / ventral striatum (wanting, motivation, addiction) · the prefrontal cortex (the goal-directed system that loses ground to habit) · the hippocampus (the other memory system — conscious/episodic vs. the basal ganglia's procedural) · the cerebellum (the co-author of motor automation) · dopamine and reward prediction error (the teaching signal) · and neuroplasticity (the substrate that lets all of this be written in the first place).
You are run by two systems: a slow one that decides what's worth doing, and a fast one that just does what you've done before — and every repetition hands more of your life from the first to the second.