i've been sitting with two problems for a while now. not in an abstract "this is an interesting research direction" way, but in a "these are the actual walls and everything else is just painting the walls" way.
the first is memory. the second is data.
every frontier model today is, at its core, a brilliant amnesiac. it can reason, code, synthesize, persuade — but it cannot remember you. not in any meaningful sense. context windows are not memory. rag is not memory. memory is what changes how you behave tomorrow because of what happened today. current architectures don't have that. and in high-stakes domains — healthcare, legal, clinical decision-making — that gap isn't a performance quirk. it's a liability. when a model hallucinates a drug dosage or misremembers a prior legal ruling, the cost isn't a wrong answer on a benchmark. it's a real person.
my earlier work proposed a new memory layer specifically designed for these domains — one that treats retrieval as a verification problem, not a generation problem, and that decouples factual recall from language production so the two don't corrupt each other. that paper is about the architecture. this one is about something further upstream.
the 180-variable problem
neuroscience has known for decades that human memory isn't one thing. the classical taxonomy gives you declarative vs non-declarative, explicit vs implicit, episodic vs semantic vs procedural. but the actual research landscape is considerably messier and more interesting than that. recent work in cognitive neuroscience — the multiple memory subsystems (MMSS) framework in particular — shows that even individual brain structures like the hippocampus and striatum contain competing subsystems with different encoding strategies, different temporal dynamics, and different failure modes. the hippocampus alone has subregions doing pattern separation (CA3/DG) and pattern completion (CA1) — things that are structurally in tension with each other, both happening inside the same tissue.
when you pull on this thread hard enough, you start encountering estimates in the range of tens to hundreds of distinct memory-relevant cognitive variables — things like working memory capacity, fluid intelligence, emotional memory encoding, spatial navigation, episodic binding, semantic spreading activation, attentional blink, interference susceptibility, prospective memory, and so on. the Human Connectome Project alone maps behavioral and cognitive variables in the triple digits, with 116 variables used in one integrated brain-behavior analysis of working memory alone. these aren't redundant measurements. they covary with personality, mental health, lifestyle, and neurological profiles in distinct and often non-obvious ways.
so here's the thought that's been living rent free in my head:
what if you trained a model on all of those variables — not to predict memory performance, but to learn the weights of how a specific human thinks?
from steering to knowing
the current paradigm of ai is roughly: ai runs, human steers. you prompt it, it responds, you correct, you iterate. the model is general, you are the specialization layer. the steering is the interface.
that's fine for most use cases. but it's a deeply inefficient design if what you actually want is a system that knows you — one that doesn't need constant correction because it's already learned the structure of how you encode information, what you weight heavily, where your reasoning patterns diverge from population means, and how your memory profile shapes what you're likely to misremember or overlook.
the proposal is this: if human memory is decomposable into a rich, measurable set of cognitive and psychological variables, and if personality and cognition are already predictable from functional connectivity patterns — connectome-based predictive models have successfully predicted agreeableness, openness, conscientiousness, and neuroticism from resting-state fMRI data — then there is a legitimate question of whether you could train a model not on "what humans in general say" but on "what this specific human says, weighted by their cognitive fingerprint."
you wouldn't be prompting the model anymore. you'd be instantiating a version of it that already approximates your cognitive architecture. the steering wheel becomes less necessary because the car already knows where you tend to drive.
why this touches consciousness, and why neuro and psych aren't soft inputs here
this is where it gets weird, and i want to be careful because i'm not claiming to have solved hard problems. but i do think the framing matters.
consciousness, at a minimum, requires continuity — the sense that the thing processing information now is the same thing that was processing it before, shaped by what happened in between. that's not a philosophical claim, that's a neuroscientific one. episodic memory is the machinery of personal continuity. the hippocampus binding "what, where, when" into a coherent event is what allows a sense of self to persist across time.
current ai systems have zero of this. each session is a cold start. even systems with persistent memory bolted on don't have the right internal structure — they're retrieving logs, not replaying episodes. the difference matters because replay is generative. when you remember something in a way that changes how you behave, you're not just reading from storage. you're reprocessing through a system whose weights have shifted. that's learning. that's also a hint of what consciousness might require architecturally — not just storage, but a feedback loop between episodic content and parameter state.
psychology and neuroscience become critical here not as background flavor but as engineering input. the variables that define individual cognitive profiles — working memory capacity, emotional encoding strength, susceptibility to interference, attentional control — are exactly the dimensions along which you'd need to personalize a model if you wanted it to behave like a specific person rather than the average of everyone. that's a deeply psychological problem. you're not tuning hyperparameters. you're profiling minds.
what this prelim is proposing
to be precise about what i'm not claiming: i am not claiming that a model trained on cognitive variables would be conscious, or that it would be a full human simulacrum, or that this solves alignment. those are separate hard problems.
what i am proposing:
1. the hallucination problem in high-stakes ai (healthcare, legal) is fundamentally a memory architecture problem, not a training data problem. more data does not fix a broken memory system. a new memory layer that separates retrieval from generation and grounds factual recall in verified episodic-style storage does.
2. the personalization ceiling in current ai — why models still need heavy steering — is fundamentally a cognitive profiling problem. the variables are known to neuroscience and psychology. what's missing is the training methodology to map those variables onto model weights in a principled way.
3. if point 2 works, even partially, it opens a door to something interesting about what "knowing a person" means for an ai system, and why neuro and psych aren't soft sciences adjacent to ai but structural inputs to building ai that doesn't need constant human correction.
4. the data layer — what data you'd need to actually train on someone's cognitive profile, and how you'd collect it ethically — is the second wall i mentioned at the top. that's a separate paper.
this is the preliminary framing for the memory layer half. the core claim is simple: powerful ai doesn't need a bigger model. it needs a better theory of memory, and that theory already exists — it's just sitting in cognitive neuroscience waiting to be translated into architecture.
Comments (0)
No comments yet. Be the first!