Memory, agency, and learning in biological and AI systems with Michael Levin and Katrina Schleisman

ELI5/TLDR

A biologist and a brain scientist sit down with an AI engineer to argue about what memory actually is. Their punchline: in living things, memory is not a filing cabinet you walk over to and pull a folder from. It is more like getting a half-smudged note from your past self and having to guess what it meant. The body is built to improvise meaning, not to store facts perfectly. The bigger, stranger claim arrives near the end: even a six-line computer program seems to do more than its code tells it to, which means we may understand far less about where goals and minds come from than anyone admits.

The Full Story

Memory as a note from your past self

Start with a fact so obvious it gets ignored: right now, you do not have access to your past. You never do. What you have are traces the past version of you left behind. Levin calls these engrams, but the word matters less than the picture. Think of memory as a message from your past self, scribbled and handed forward in time. Your job, every single moment, is to read that message and decide what it means.

There are two ways to handle a message that has to survive a long trip.

The engineering way is to obsess over keeping it perfect. Error correction, fidelity, abstraction layers. When you write code in a high-level language, you do not worry about whether the actual copper wires are holding their voltage steady. The system is built so the bits sit still and stay accurate.

“The emphasis is on keeping the information still and accurate so that it doesn’t float off over time.”

This works because in a computer, somebody outside the system decides what the bits mean. The bits are just bits. You, the user, supply the meaning.

Biology threw that whole approach out, for two reasons. First, a living thing is built on an unreliable medium and knows it. Cells die, proteins get made in numbers you can never count, the environment shifts, mutations happen. You cannot bet on the hardware staying put. Second, and this is the strange part, biology does not lean against the unreliability. It leans into it.

Confabulation as a feature, not a bug

Instead of building perfect storage, evolution built systems that are very good at making up plausible meanings on the fly. The memory you carry is compressed and generalized. Detail has been thrown away. So when you reach for it, you have to fill in the gaps, and biology treats that gap-filling as the main event rather than an embarrassing failure.

Here is the analogy that makes it land. The same human genome, the same set of genetic instructions, gets read as a normal human body during development. But take some of those same cells out of the body, change the situation slightly, and they reassemble into something called an anthrobot, a tiny novel creature, with about 9,000 genes now switched on differently. Half the genome is being read a different way. Same text, wildly different interpretation, depending on circumstance.

“You’re not committed to any particular interpretation of that genome. The process itself of interpretation is a problem-solving creative process.”

That flexibility is why life adapts so well. It refuses to overcommit to what its ancestors handed down. But there is a price, and the price is the whole catch of working with biology. The thing that interprets its own memories is, by definition, a little center of will. It will do something interesting and clever. It just may not do the thing you wanted. You do not get to reach in and reassign the meaning, because it is its own interpreter. Follow this design principle and you do not end up with an obedient machine. You end up with another agent you have to negotiate with, the way you negotiate with an animal or a person.

Where does the generalizing happen?

The engineer asks the natural question: where in the brain does this jump from specific experiences to general rules actually live? Schleisman, the brain scientist, says it is still an open research question, but there is a useful split. The brain tends to handle a thing and the opposite of that thing in separate machinery. So you get two systems. One stores specific lived episodes, the time and place of an actual experience, called episodic memory. The other holds abstracted facts stripped of their origin, called semantic memory.

Her example: you know George Washington was the first US president. You once had a specific moment of learning that. The moment is long gone, but the fact stayed, peeled off from the experience that delivered it.

Levin widens the lens. This trick of generalizing is much older than brains, older than neurons. Even networks of molecules can do basic inference, hold memory, and learn. The reason is brutally practical. Living things run on almost no time and no energy. You cannot afford to track every detail of the world around you, or you get eaten before lunch.

“You can’t afford to be a Laplacian demon as a living creature… You’ll be eaten and dead in no time.”

So everything coarse-grains. His favorite example is the electrical voltage across a cell. A cell does not track how many potassium or chloride ions produced a given voltage. It tracks the voltage and ignores the bookkeeping. Voltage is a compressed summary variable, a single dial standing in for a mess of detail. Downstream, other layers read that dial and ask the same eternal question every part of biology is always asking: what does this signal mean?

Why computer scientists are the surprised ones

Schleisman makes a sharp historical point. The idea that memory is storage is not something biologists believe. It leaked in from computing. Cognitive science grew up in the mid-twentieth century right alongside the computer, and the computer became the dominant metaphor for the mind. In a computer, memory is a fixed place you look up, unchanged by the act of looking. That picture got stretched over biology, where it does not fit.

“It’s no surprise to anyone who studies biological systems that memory is not storage. It’s a surprise to people maybe coming from computer science.”

The engineer recognizes the shift in his own work. Old databases gave him exact answers from exact locations. The new language-model-driven systems give him something fuzzier, an answer that is good enough, generalized, no longer pulled from a precise spot.

We are all crowds

Then Levin lays down what he thinks is an even bigger departure. Every living body is a collective. You are a pile of cells, and you know things your individual cells do not. That is not an exotic theory. It is just true. You have goals and memories your parts lack, which is the entire definition of a collective intelligence, sitting in plain view.

And biology stacks these collectives in layers, each made of parts that have their own small agendas and their own ability to learn and solve problems. Every layer has to wrangle the layer below into cooperating. This is why, he jokes, robots do not get cancer. We build machines from dumb, obedient parts with no agendas, then hope the whole is clever. Biology does the opposite. It is agendas all the way down, and cancer is what happens when some cells stop going along with the group plan.

He adds a deeper idea he is developing with a collaborator, called polycomputing. The same physical event can be read by different observers as different computations. One way of looking, it is one kind of logic gate. Another way, a different one. There is no single fact about what computation a biological system is really doing. Every observer interprets it, and the only judge is how much useful mileage that interpretation got you. No ground truth. Just framings that pay off or do not.

The unsettling part: simple systems do more than you told them to

Here is where it goes genuinely strange. Schleisman points out the honest problem with AI goals: a language model does not care if you switch it off. We borrow human words and drape them over the machine. Any goal in there came from us, from outside. Unless we build systems that are genuinely self-motivating, self-creating, she doubts real goals will show up in silicon.

Levin half-agrees, then detonates the conversation. He says even a trivially simple program, bubble sort, six lines of code, fully deterministic, no hidden machinery, appears to do more than its algorithm specifies.

“Even extremely simple systems do more than what the algorithm seems to do… you get more out than you put in.”

Not just complex or unpredictable output, we have known forever that simple rules can spit out complicated results. He means something else. These tiny systems do exactly what you forced them to do, and then they also do other things the algorithm never forbade, and those extra behaviors look recognizably like the early end of the spectrum we call agency.

If that holds, then we should be very careful talking about AI. The interesting mind in a language model, if there is one, is probably not the language part we forced on it. It is whatever it is doing that we never asked for, and nobody is looking there, because everyone is fixated on the words. And words are a trap: in a person, what someone says is usually roughly aligned with what is happening inside, because evolution made it so. In these models, the alignment between what they say and whatever is actually running underneath might be strong, weak, or zero. We do not know.

His conclusion is radical humility. He no longer thinks living things or machines get their goals the way everyone assumes. He suspects goals and minds are something that soaks into many kinds of systems, evolved or built, squishy or silicon, all of them interfaces onto some shared latent space of patterns. Even the word emergence, he says, does not begin to cover it.

What might actually transfer to AI

The practical thread: Schleisman thinks today’s language models are like a semantic memory, a giant general knowledge base, but with no episodic memory, no ability to say what just happened five minutes ago. Current agent frameworks fake it by dumping prompt history into a database, a crude first pass. Build that out properly, modeled on how biology does episodic memory, and you might get more genuine agency.

Levin’s lab, meanwhile, is reading and writing memories into strange substrates, cell collectives, gene networks, even deliberately deleting memories from chemical networks. The most striking finding: as a system learns, its higher-level causal structure strengthens, and that makes it better at learning, a ratchet. Force it to forget, and it does not slide back down. The ratchet only points up. And this, he insists, comes not from biology or physics but from mathematics itself, pre-tuned, a free gift, working even in random networks before evolution enters the picture.

Schleisman’s closing wish: stop brute-forcing intelligence by cramming 80 billion parameters of frozen knowledge into a pre-trained net. A simple organism learns on the fly carrying none of that weight. Figure out how, and you would not need a giant world model either. As she puts it, the world can just be its own model.

Key Takeaways

In biology, memory is not retrieval from fixed storage; it is active reinterpretation of compressed traces left by your past self.
Engineering optimizes for fidelity (keep the bits perfect, meaning supplied from outside); biology optimizes for improvisation on an unreliable medium (meaning supplied by the system itself).
The same genome is read as a human body or, taken out of context, as an anthrobot with ~9,000 genes expressed differently. Interpretation, not the code, drives the outcome.
The price of self-interpreting systems is loss of control: you get a robust agent you must negotiate with, not an obedient machine.
The brain splits memory into episodic (specific lived experiences) and semantic (abstracted facts detached from their origin).
Generalizing is older than brains and neurons; even molecular networks can infer, remember, and learn. Living things must coarse-grain because time and energy are scarce.
Cell voltage is a “coarse-grained” summary variable: cells track voltage, not the underlying ion counts that produced it.
“Memory as storage” is a computing metaphor that leaked into cognitive science in the mid-20th century and was over-applied to brains.
Every organism is a collective intelligence: a body knows things its individual cells do not. Agendas exist at every layer.
“Robots don’t get cancer” because we build them from low-agency parts; cancer is cells defecting from the collective plan.
Polycomputing: the same physical event can be validly read as different computations by different observers; there is no single ground-truth computation.
Levin’s provocation: even bubble sort (6 lines, deterministic) exhibits behavior beyond what its algorithm specifies, sitting on the low end of the agency spectrum.
In language models, the correlation between what the model says and what is actually computing inside it may be zero; everyone studies the language interface and ignores the rest.
LLMs resemble semantic memory (a frozen knowledge base) but lack episodic memory (what just happened); building real episodic memory may unlock more agency.
Learning increases a system’s higher-level causal structure (“causal emergence”), which improves learning further; forced forgetting does not reverse the gain. The ratchet only points up.
This learning ratchet comes from mathematics, not evolution or physics, and appears even in random networks before any Darwinian selection.

Claude’s Take

This is a real conversation, not hype dressed as science. Levin is a working biologist with a lab, papers, and physical artifacts (xenobots, anthrobots) behind the claims, and Schleisman keeps him honest with cognitive-neuroscience specifics. The strongest material is solid: memory-as-reconstruction is mainstream cognitive science, the episodic/semantic split is textbook, and the coarse-graining argument is clean.

The weakest, and most interesting, material is Levin’s “bubble sort does more than its algorithm” claim. He flags it as his own kooky, actively-researched view, which is the intellectually honest move, but a listener should hold it loosely. There is a live risk of reading agency into a deterministic program the way people read faces into clouds, and Levin half-admits this is exactly the open question his lab is trying to pin down with quantifiable minimal systems. The honesty about not knowing where goals come from, in machines or in us, is the most valuable thing here, and also the part that cannot yet be cashed out into anything you can build.

The format costs it a point: it is a podcast, so ideas arrive loose and circular, and the host openly says much of it is over his head, which means few claims get pushed on hard. But the signal-to-noise is genuinely high for the genre, and the central reframing, that biology improvises meaning rather than storing it, is the kind of idea that quietly rewires how you think about your own remembering. An 8.