Neural Atlas · Health
Care for the original. Use the copy well.
For each region, two sides of the same parallel: what the science says about keeping the brain healthy, and the concrete best practices for getting the most out of its AI counterpart. The biological notes are anchored to meta-analyses and peer-reviewed studies; the AI notes to official documentation and research.
An accessible summary of the scientific literature, not medical advice: it is no substitute for a physician or qualified health professional. The link between an intervention and a single brain region is often at the level of function, not an isolated structure.
Hippocampus
For the brain
- Aerobic exerciseRCT
In a one-year trial of 120 older adults, aerobic exercise grew the anterior hippocampus ~2% (reversing one to two years of age-related loss) while controls shrank; gains tracked serum BDNF.
Erickson et al.: Exercise increases hippocampus size, PNAS (2011) - Deep sleepMeta-analysis
In NREM sleep the hippocampus replays the day and hands memories to the cortex; a 2025 meta-analysis found slow-oscillation–spindle coupling the strongest predictor of next-day retention.
Slow oscillation–spindle coupling & memory: Bayesian meta-analysis, eLife (2025) - Omega-3 (EPA/DHA)Cohort
In 1,111 women, a higher omega-3 index predicted ~50 mm³ more hippocampal volume eight years later.
Pottala et al.: Omega-3 index & hippocampal volume (WHIMS-MRI), Neurology (2014) - Manage chronic stressCohort
Chronically high cortisol suppresses hippocampal neurogenesis and tracks volume loss in depression: effects that prevent more than they restore.
Stress, depression & hippocampal morphology, Biological Psychiatry (2018)
Volume is a stand-in for memory, and the strongest data come from older or clinical groups; omega-3 and stress findings are observational. These attenuate decline rather than reverse damage.
For the AI counterpart
- Externalize into artifacts
A model's working memory vanishes at each reset; write decisions and state into durable files it re-reads, the way the hippocampus offloads to cortex.
Anthropic: Effective context engineering for AI agents - Start fresh context windows
Accuracy degrades as a window fills: “context rot” sets in well before the nominal limit, so begin a clean session for each new task.
Chroma: Context Rot, how input length degrades LLM performance (2025) - Front-load what matters
Models retrieve best from the start and end of the context and worst from the middle; put critical information at the edges.
Liu et al.: Lost in the middle, arXiv:2307.03172 - Use a memory tool or RAG
For durable, cross-session recall, a dedicated memory store or retrieval system beats stuffing the whole history into every prompt.
Anthropic: Memory tool (Claude docs)
Amygdala
For the brain
- Mindfulness (MBSR)Longitudinal
After eight weeks of MBSR, drops in perceived stress tracked measurable decreases in right basolateral amygdala gray-matter density.
Hölzel et al.: Stress reduction & amygdala structure (2010) - Protect your sleepReview
One sleepless night raised amygdala reactivity to negative images by ~60% and uncoupled it from prefrontal control.
Krause et al.: The sleep-deprived human brain, Nat Rev Neurosci (2017) - Meditation that transfersRCT
Eight weeks of mindful-attention training lowered amygdala response to emotional images even in an ordinary, non-meditative state.
Desbordes et al.: Meditation & amygdala response in an ordinary state (2012) - Social supportObservational
Perceived social support buffered the link between threat-related amygdala reactivity and trait anxiety.
Hyde et al.: Social support, amygdala reactivity & anxiety (2010)
Samples are small and some structural findings did not replicate in larger preregistered trials; sleep studies use near-total deprivation. Effects are modest and no substitute for clinical care.
For the AI counterpart
- Keep a human in the loop
Before consequential or irreversible actions, retain the ability to stop and approve; capable models still misread intent and act on bad premises.
Anthropic: A framework for safe and trustworthy agents (2025) - Verify, don't over-trust
Fluent, confident output is still sometimes wrong; review anything high-stakes rather than taking it at face value.
OpenAI: Safety best practices (API docs) - Set explicit guardrails
Stating your values, scope and prohibited outputs in the system prompt is the user-level analogue of Constitutional AI: top-down rules that shape behavior.
Anthropic: Constitutional AI, harmlessness from AI feedback - Extra scrutiny in high-stakes domains
In legal, medical or financial use, govern AI with human oversight proportional to the impact: the spirit of the NIST AI Risk Management Framework.
NIST: AI Risk Management Framework (AI RMF 1.0)
Cerebellum
For the brain
- Moderate aerobic exerciseMeta-analysis (animal)
A 2025 meta-analysis in animal models found moderate exercise cut cerebellar oxidative damage and protected Purkinje cells, while excessive training raised it: a dose-dependent window.
Exercise & cerebellar neuroplasticity: meta-analysis (animal models), PLoS One (2025) - Deliberate motor practiceNeuroimaging
A new motor skill leans on the cerebellum while you learn it, then migrates to striatal-cortical circuits as it becomes automatic.
Doyon et al.: Cerebellar contributions to motor sequence learning, PNAS (2002) - Balance & coordination trainingMeta-analysis
16 RCTs (~2,960 adults over 65) found targeted exercise improved static and dynamic balance and reduced falls.
Exercise, balance & fall prevention: meta-analysis of 16 RCTs (2020) - Limit heavy alcoholAutopsy cohort
The cerebellum is selectively alcohol-vulnerable: dose-dependent Purkinje-cell loss runs ~15% at 41–80 g/day and ~33% above.
Alcohol & cerebellar Purkinje-cell loss, BMJ (1994)
The exercise meta-analysis is in rodents, balance trials use mixed protocols that can't isolate the cerebellum, and the alcohol data are observational. Pathway specifics come from imaging and animal models.
For the AI counterpart
- Give clear, distinct tools
Agents act through tool descriptions; well-named, well-documented, non-overlapping tools are the single biggest lever on reliability.
Anthropic: Writing effective tools for AI agents - Build verification loops
Let the agent check its output against explicit rules before finalizing; caught early, errors don't compound across a long task.
Anthropic: Building effective agents - Decompose long tasks
Break a long-horizon goal into discrete, verifiable steps and validate each before the next, the way the cerebellum chains sub-movements.
Anthropic: Building effective agents
Prefrontal Cortex
For the brain
- Aerobic exerciseMeta-analysis
Across RCTs in sedentary adults, regular aerobic activity yields a small but real gain in executive function (g≈0.37), via raised BDNF and cerebral blood flow.
Aerobic activity & executive function: systematic review & meta-analysis (2021) - Mindfulness trainingMeta-analysis
111 RCTs (9,538 people) show improved inhibition (g=0.64), working memory and executive attention: all prefrontal functions.
Mindfulness & cognition: meta-analysis of 111 RCTs, Health Psychology Review (2023) - Protect your sleepReview
Sleep deprivation reliably depresses dorsolateral prefrontal activity during attention and working-memory tasks; the PFC is among the regions hit hardest by sleep loss.
Krause et al.: The sleep-deprived human brain, Nat Rev Neurosci (2017) - Manage chronic stressCohort
Sustained cortisol elevation disrupts prefrontal circuits; higher cortisol reactivity is linked to lower executive function through prefrontal structural change.
Cortisol reactivity & prefrontal executive function (2021)
Effects are small-to-moderate and strongest in older or clinical groups; the stress link rests largely on observational and animal data. None of this is a clinical prescription.
For the AI counterpart
- Turn on extended thinking
For multi-step problems, let the model reason before answering; extended/“thinking” modes allocate a reasoning budget that lifts planning and analysis.
Anthropic: Building with extended thinking (Claude docs) - State the goal and success criteria
An underspecified goal makes a reasoning model optimize the wrong target; spell out the end state and any constraints up front.
OpenAI: Reasoning best practices (API docs) - Give it room to think
Set a generous output budget and a higher effort level; too tight a limit cuts the model off before it finishes reasoning.
Anthropic: Extended thinking tips (Claude docs) - Few-shot for novel reasoning
When a problem needs an unusual reasoning pattern, a couple of worked examples (chain-of-thought) elicit it better than instructions.
Wei et al.: Chain-of-thought prompting, arXiv:2201.11903
Basal Ganglia
For the brain
- Aerobic exerciseMeta-analysis
Exercise supports striatal dopamine function; across 8 cohorts (544,336 people) the most active had 29% lower Parkinson's risk.
Physical activity & Parkinson's risk: meta-analysis, JAMA Network Open (2018) - Context-consistent repetitionField study
Habits form by repeating a behavior in a stable context until the striatum automates it: a field study found this took a median ~66 days, and one missed day didn't break the streak.
Lally et al.: How are habits formed, Eur. J. Social Psychology (2010) - Deliberate, spaced practiceReview
Dopamine-dependent plasticity at cortico-striatal synapses is what turns an effortful skill into an automatic one.
Tian & Chen: Mechanisms of motor skill learning, Frontiers in Neurology (2021) - Protect the reward systemReview (animal)
Engineered supernormal rewards (addictive substances, ultra-processed food) drive excess striatal dopamine that erodes flexible action selection.
Yager et al.: The striatum in drug addiction, Neuroscience (2015)
The Parkinson's link is observational (reverse causation can't be fully ruled out), and habit and reward findings lean on self-report and animal models. Genetics and age strongly moderate all of it.
For the AI counterpart
- Prompt before you fine-tune
Prompt engineering is often all you need; exhaust it (with a real evaluation baseline) before paying for fine-tuning.
OpenAI: Optimizing LLM accuracy (platform docs) - Build repeatable workflows
Templated prompts that fix role, objective and format encode a good “habit” into the workflow and cut output variance: the software analogue of striatal chunking.
Anthropic: Building effective agents - Run an evals feedback loop
Evaluations are the highest-bandwidth signal for improvement: they surface failures before production and prevent regressions, mirroring reward-prediction-error learning.
Anthropic: Demystifying evals for AI agents
Thalamus
For the brain
- Protect your sleepfMRI study
Sleep spindles are generated by the thalamic reticular nucleus; 36 hours without sleep measurably weakens thalamus–cortex connectivity and dulls arousal and attention.
Sleep deprivation & thalamocortical connectivity, PLoS ONE (2013) - Cardiovascular healthCohort
The thalamus is among the most stroke-vulnerable deep structures; controlling blood pressure and vascular risk protects the nuclei that gate attention.
White-matter damage & thalamic atrophy in small-vessel disease, Alzheimer's & Dementia (2025) - Aerobic fitnessCross-sectional
Higher cardiorespiratory fitness is associated with altered thalamic perfusion and preserved cerebrovascular reactivity, though the evidence here is only cross-sectional.
Aerobic fitness & cerebrovascular function, Frontiers in Physiology (2020)
Region-specific evidence for the thalamus is genuinely thin: no trial has tested whether any behavior selectively improves thalamic function in healthy people. These are general brain- and vascular-health measures whose benefits plausibly reach thalamic circuits; the vascular link is the best-evidenced.
For the AI counterpart
- Give clean inputs per modality
Clear, legible, lightly-compressed media on each channel; the model hallucinates on low-quality or tiny inputs.
Anthropic: Vision (Claude docs) - Point to what matters
Tell the model which image, region or modality to focus on, and have it describe each input before reasoning across them.
Google: Design multimodal prompts (Vertex AI docs) - Mind the text bias
Multimodal models systematically underweight images and audio relative to text; don't assume your visual input was fully used without grounding it in words.
Wu et al.: Text dominance in multimodal LLMs, arXiv:2508.10552
Visual Cortex
For the brain
- Aerobic exerciseNarrative review
Raises BDNF and, in animal models, lowers the inhibitory threshold for plasticity in primary visual cortex. Human evidence is suggestive but less consistent.
Aerobic exercise & the human visual cortex: Neural Plasticity (2022) - Lutein & zeaxanthinRCT
These macular carotenoids concentrate in retinal and occipital tissue; six months of supplementation improved visual memory and learning in adults with mild cognitive complaints.
Lopresti et al.: Lutein & zeaxanthin RCT, Frontiers in Nutrition (2022) - Time outdoorsMeta-analysis
Daily outdoor light protects the developing eye: a synthesis of seven reviews (~64,000 children) cut myopia incidence by 24–46%, preserving the input the visual cortex is built on.
Time outdoors & myopia: overview of systematic reviews, OPO (2022) - Perceptual learningReview
Repeated practice on hard, near-threshold visual tasks drives plasticity in early visual areas; the adult visual system keeps real adaptive capacity well into rehabilitation.
Legge & Chung: Low vision & plasticity, Annual Review of Vision Science (2016)
Much of the strongest evidence (outdoor light, carotenoids) acts on the retina, not primary visual cortex itself. Only perceptual learning has direct human V1 imaging support, and no large trial shows selective V1 gains in healthy adults.
For the AI counterpart
- Send sharp, high-resolution images
Heavy compression and tiny or rotated images cause errors; provide the clearest version, and place the image before your text.
Anthropic: Vision (Claude docs) - Ask narrow, specific questions
Vision models struggle with precise localization and exact counts; a scoped question (“how many in the top-left?”) beats “what do you see?”.
OpenAI: Images and vision (API docs) - Verify OCR, counts and positions
Spatial and counting hallucinations are systemic across vision models, not a quirk of one, check anything load-bearing.
HallusionBench: vision-language hallucination, arXiv:2310.14566
Broca & Wernicke
For the brain
- BilingualismReview
Managing two languages builds cognitive reserve: bilinguals show dementia symptoms about 4.5 years later than monolinguals, a delay in symptoms, not prevention of disease.
Berkes & Bialystok: Bilingualism as cognitive reserve (2022) - Reading & verbal engagementMeta-analysis
Reading reliably recruits left-hemisphere language regions including Broca's area; 163 neuroimaging studies confirm it as a primary workout for the language network.
The reading brain: meta-analysis of 163 studies (2025) - Social connectionMeta-analysis
Across 38,614 adults in 13 cohorts, living with others, regular contact and the absence of loneliness predicted slower cognitive decline.
Social connection & cognition: IPD meta-analysis, Lancet Healthy Longevity (2022) - Musical trainingMeta-analysis
Music and language share syntactic processing in the inferior frontal gyrus; 84 studies find musicians engage these overlapping regions more strongly.
Musical expertise: ALE meta-analysis, Scientific Reports (2022)
These build reserve across a distributed bilateral language network, not Broca's or Wernicke's area in isolation, and they delay symptoms rather than stop disease. Bilingualism evidence in particular is heterogeneous.
For the AI counterpart
- Write explicit instructions
LLMs follow instructions, not intentions; state the task, the constraints and the output format directly instead of leaving the model to guess.
Anthropic: Prompting best practices (Claude docs) - Give a role and context
Framing the domain, audience and tone narrows the model's uncertainty and calibrates vocabulary and register to your use case.
OpenAI: Prompt engineering (API docs) - Provide a few examples
Two or three input–output examples communicate format and reasoning pattern better than description alone: the foundation of few-shot prompting.
Brown et al.: Language models are few-shot learners, arXiv:2005.14165