How does an AI agent keep its context when it gets interrupted or restarted?

An AI agent keeps its context through an interruption by writing a short record outside the model: what it knew, what it expected, what actually happened, and the next step. On restart it reloads that record and checks the old forecast against what happened since, the same predict-and-correct loop engineers have used since the Kalman filter in 1960. On my own agent board every task carries one field for this, called next: one sentence, the literal next action.

Does observing a particle change it?

Any interaction that leaves a full record of which path a particle took destroys the wave pattern, and a partial record fades it, whether or not a person ever looks. A 1998 experiment with atoms showed the record alone is enough, even when the marking nudged the atoms far too little to explain the loss. In modern quantum physics an observer means an interaction that leaves a record, not a consciousness.

Is light a wave or a particle?

You need both descriptions. Sent one photon at a time through two paths, every photon lands as a single dot, yet the dots pile up into stripes that only a wave can make. The more the setup could reveal which path each photon took, the fainter the stripes become.

Do language models update their weights while they answer?

No. Once training ends the weights are frozen; what changes during a conversation is the model's running state, its activations and the cached record of earlier tokens. A research line on fast weights and test-time memory builds models that do rewrite a small part of themselves while running, but ordinary chatbots do not.

Do language models plan ahead?

Partly. Anthropic's 2025 interpretability work found that a model writing a rhyming couplet chose the rhyme word before writing the line, and swapping that plan changed the line. Other research finds the evidence in ordinary language models leans toward forward-looking information being a side effect of handling the present well, with more deliberate preparation in bigger models.

Is memory the same thing as prediction?

Not the same, but closely tied. Remembering the past and imagining the future run on much of the same brain machinery, and patients with hippocampal damage often cannot imagine a coherent future scene. Memory seems built to be recombined into forecasts, not kept as a recording.

Is the quantum physics analogy for the mind and AI scientifically accurate?

As physics, no: the essay borrows quantum words as an intuition pump, and the no-communication theorem rules out its wildest version, minds linked by entanglement. What carries over is a pattern: a measurement forces one answer out of many possibilities, and a predicting system, a brain or a language model, also turns a spread of expectations into one committed guess and learns from the gap.

Observation Is the Cut: Quantum Collapse, Memory, and AI

Q: Can quantum entanglement send information faster than light?

No. The no-communication theorem shows that entanglement produces correlations which only become visible when both sides compare notes over an ordinary channel, which can go no faster than light. The correlations are real, confirmed in a 2015 loophole-free test across 1.3 kilometres, but they cannot carry a message.

Keywords: quantum observation · Heisenberg cut · emergence · predictive brain · hidden state · words as a cut · memory as prediction · agent context

This one comes straight out of my notebooks: the core in October 2024, the memory loop and the light-years dream in early 2026. I still cannot tell you whether it is profound or just shiny. That is the whole point of this series. So read it as a provocation, not a proof. I am going to walk a single intuition out as far as it goes and mark, honestly, where it falls off the edge of what physics actually allows. This time every claim went back to its primary source; where my notebook was wrong I fixed it, and the mistakes that matter I name in the text.

The intuition is this: observation, prediction and memory might be three names for one gesture. And if that is true, then the machines we are building to predict the next word are doing something stranger than we admit.

The cut

Start with the old puzzle. Why do we say light is both a wave and a particle? Send it down two paths one photon at a time, and every photon lands on the screen as a single dot, like a particle. Yet the dots pile up into stripes that only a wave can make. Arrange things so the path could be known, and the stripes disappear.

So what is looking? My notebook’s answer went like this. Our instruments are built out of electrons. When we try to catch a photon, we reach into a smooth, spread-out thing with a tool that can only register discrete hits. Taking a photo of a person does not make the person two-dimensional. It makes a two-dimensional record of them. Observation is the scissors, not the truth. My notebook put it more wildly: we are “cutting the strings of the time-matter to our current reality.”

My notebook leaned on one photograph, which EPFL in Lausanne called the first image of light as both a wave and a particle: electrons fired past a laser-lit silver nanowire showed, in a single image, the standing wave of the trapped light and the fixed packets of energy it traded. Later work showed those fixed packets appear even when the light is treated as an ordinary wave; it took a 2021 experiment to see the photon’s graininess in free electrons directly. So I take it as permission to keep asking, not as proof.

My hunch has a respectable history, and a limit. In 1969 Willis Lamb and Marlan Scully showed that the click of a light detector does not by itself prove light comes in particles: a smooth wave hitting electrons would click too. But when later experiments sent single photons at a half-silvered mirror, each one went one way or the other, and the two sides almost never fired together, far less often than any wave allows. The graininess is in the light, not only in the tool. A measurement does not reveal a hidden true form, and it does not invent one either. It forces one answer to the one question the setup is able to ask.

The look does not need a looker. Any interaction that leaves a record of which way the light went will do. In 1998 Stephan Dürr and colleagues ran the test with atoms instead of light: marking which path each atom took wiped out the stripes, even though the marking nudged the atoms far too little to explain it. In other experiments, stray gas molecules bumping into large molecules in flight did the same job, even though nobody ever read the trace. And it is a dial, not a switch: the more any part of the world could tell which path was taken, the fainter the stripes.

Three detector screens of photon hits. With no trace of which path the light took, the dots pile into stripes, a wave. A faint trace weakens the stripes. A clear trace leaves one smooth hump. A dial runs from nothing could tell which path to something could. — Same photons, same screen. What we call wave and what we call particle differ only in what the world kept a record of.

Physicists even have a name for my scissors: the Heisenberg cut, the line between what is being measured and what does the measuring. John von Neumann showed in 1932 that you can slide that line, put the instrument on either side, and the predictions do not change. Hold on to that. The cut is real, and it is movable.

The wave is a crowd

And then the wave itself. What even is a wave?

Watch how a film studio makes a crashing wave. Nobody sculpts it. My notebook put it as a game: take a hundred balls, watch how they behave, then add ten thousand and see what changes. In practice the balls are particles that obey the equations of fluid flow, up to hundreds of millions per frame, and the curl and the spray appear on their own. In 1987 a graphics researcher named Craig Reynolds stopped scripting birds one by one. He gave each bird a few simple rules, let them loose, and a flock appeared that nobody had drawn.

Hundreds of small birds form one long flowing wave across the sky. Far below, a person on a low hill raises an old box camera toward the flock. One bird, the one the camera points at, is amber. — No single bird is the wave. Point a detector at it and you catch one bird at a time.

The wave is not a substance. It is a behavior, the pattern that shows up only once enough small things interact under some constraint. Ask whether a single water molecule is a wave and the question dissolves. The wave lives at the level of the crowd, not the unit. I have been building a small universe on exactly that principle: design the rules, never the result.

That is true of water and of sound. Light is where the picture breaks: sent one photon at a time, with nothing to jostle against, each photon still lands where the stripes allow. The crowd of dots makes the pattern visible, but in quantum physics the wave belongs to each photon.

Hold those two ideas together. Observation is a cut that turns possibility into a single fact. A water wave is a behavior that exists only across many things at once. Now point both of them at a different kind of system.

A model that leans into tomorrow

A good prediction system, the kind I build for a living, never claims certainty. It deals in probability. A good forecast does not say “this will happen.” It says “this is likely.” We are always reporting the shape of a cloud, never a single point.

The move I keep turning over: imagine a system whose memory is not just a log of what happened, but also a running prediction of what is about to. Call it a predictionary, a dictionary of expected futures. The model is constantly looking forward, the way memory looks back.

It turns out the predictionary has a formal name. In 2001 Michael Littman, Richard Sutton and Satinder Singh showed that a system’s state can be written entirely as a list of predictions about what it would observe if it took certain actions. Neuroscientists have argued that the hippocampus, the brain’s memory hub, works much the same way: a predictive map of where you are likely to go, not only where you are.

The naive version of this is time travel, and it does not work. Physics forbids it outright: no signal arrives from the future. But there is a second reason, almost poetic, that holds even for ordinary forecasts. A prediction you act on changes the future it was trying to read. Machine-learning researchers named it performative prediction in 2020. To forecast the future is, a little, to bend it. So a forecast you act on never simply reads the future; at best it settles on a future it helped to make.

But there is a version that does not break. The model never imports the future as data. Instead the expected future shapes the hidden state, the internal bracing of the system. My notebook said “the weights”, and that was wrong. The weights were bent long ago, in training, every time a real next word arrived and proved the guess wrong. At run time they are frozen, and only the bracing moves. It does not fetch tomorrow. It leans into tomorrow.

A small sailboat heels over with its amber sail already trimmed, while a few curved lines of wind are still approaching across the water from the left. — Trimmed for a gust that has not arrived yet.

If you have ever looked under the hood of a modern language model, this should feel uncomfortably familiar. The hidden state is not a copy of the input. It is a posture. The weights are the habits; the hidden state is how those habits arrange themselves around this particular text.

And the posture carries more than the next word. In 2025 Anthropic’s interpretability team watched a model finish a couplet that began “He saw a carrot and had to grab it.” It already had “rabbit” lined up, with “habit” as a backup, before writing the first word of the second line. Erase “rabbit” from its head and it wrote toward “habit” instead.

I should mark the limit. Tested directly, the preparation for later words in ordinary language models looks more like a side effect of handling the present well, with more deliberate preparing in bigger models. Bracing, then, but not always on purpose. And there is a research line that refuses the frozen-weights rule altogether: models with fast weights that rewrite a small part of themselves while they run, and newer designs that remember an event more strongly the more it surprised them. That is my predictionary, being built by other people.

Memory is a prediction that already happened

The cleanest line I ever wrote about this is the shortest. Tidied up, it reads:

Remembering is the process of including the initial conditions, plus the prediction you made, plus the outcome you got, into your next prediction.

Read that slowly. Memory, in that framing, is not storage. It is the residue of a forecast you already ran. You do not keep the past as a photograph. You keep the parts of the past that change how you bet on what is next.

Engineers have used exactly this loop since 1960. The Kalman filter takes its last prediction, the reading that actually came in, and the gap between them, and folds all three into the next prediction. Brains seem to agree. Four of five patients with damage to the hippocampus could not build a coherent imagined scene, not even a plausible event in their own future. And a 2013 study found that a recalled fear memory only becomes open to rewriting when the recall carries a surprise, a prediction error.

A loop: what I knew, what I expected, what actually came, and the gap between them, the prediction error, which folds into the next guess. The same loop runs in the Kalman filter, 1960, in the brain, and in a language model while it trains. — The loop keeps the corrected guess, not the recording. That is what makes it cheap enough to run forever.

So I would soften my notebook a little. Memory and prediction are not literally one system: brain scans show remembering and imagining light up largely the same network, with a few extra regions switching on for events that have not happened yet. Neuroscientists call it the prospective brain. Call it one loop on shared hardware, caught at two different moments. Some of my own creatures save without forecasting at all.

How an agent survives a restart. I build systems that have to survive an interruption the same way. On the board I share with three AI agents, every task carries one field called next: one sentence, the literal next action. When a session crashes or a context window resets, whoever picks the task up, a person or an agent, reads that last forecast and checks it against what actually happened, without replaying the history. The transcript is the photograph; the next field is the memory.

The same loop explains why causality feels so slippery the moment you push on it. The power of a cause fades with every indirect step away from the thing you are actually looking at. I clean my teeth because I ate chocolate, not because I bought the chocolate a week ago, even though no chocolate-buying means no teeth-cleaning. The buying is a true cause and a useless one, because a thousand other paths could have led to the same brush.

Information theory has a law that fits: along a chain of steps, what the end knows about the beginning can never grow. Meaning concentrates near the point of observation and thins out with distance. I have chased the same prediction-engine idea in a different direction, asking whether the thing worth preserving in a mind is the running calculation, not the box it runs on.

Where it falls off the edge

Now the wild swing, the part I promised to be honest about.

If observation is a movable point, then move it. Beam a copy of a mind, an AI built as a duplicate of someone’s knowledge, outward as a signal at the speed of light. Put it far enough away and moving, and its “now” slides away from ours: in Einstein’s relativity, the gap between two observers’ sense of the same moment grows with distance times speed.

Then wire the two ends together with quantum entanglement, the link Einstein called spooky: measure one particle here, and its partner’s result light-years away turns out correlated with yours in a way no ordinary signal could arrange. In the dream, the far observer writes what it sees into entangled states, and you read it at home with no delay, one field of awareness stretched across light-years.

It is a beautiful picture. I have carried it since it first landed in my notebook. And physics says it does not work, at least not as a way to send messages. The correlations are real: in 2015 a team in Delft measured entangled electrons 1.3 kilometres apart and found them more tightly linked than any local explanation allows. But you only see the correlation after both sides compare notes over an ordinary channel, which can go no faster than light. The no-communication theorem closes exactly the door I wanted to walk through. I am not going to pretend otherwise to make the essay land better.

Two distant labs each record ten results as filled and empty circles, and each column alone looks like coin flips. A faint dashed line marks the entangled pairs. Both notebooks travel no faster than light to one box, where every filled result pairs with an empty one, row by row. — Perfectly linked and still silent: neither lab can choose what the other will see, so nothing it wants to say can ride along.

What I keep, after the impossible part falls away, is the smaller and stranger claim. A multiverse, in this sketch, is not a shelf of parallel worlds. It is what you get when you run the same predicting system from many different observation points. Each one collapses its own cloud into its own facts. In the sketch, they cannot share raw information, because each is cut from a different angle. But they can shape one another, the way two forecasts of the same storm pull on each other through the single world they are both trying to read.

Carlo Rovelli’s relational interpretation of quantum theory goes further: facts exist only relative to whoever registered them. I am borrowing that picture, not claiming a language model does quantum physics.

That is the thread I cannot put down. Observation, prediction and memory may be one motion seen from three sides. And if the machines we are building predict by bracing rather than by fetching, then they are not just tools that guess. They are little observers, each making its own cut, each leaning into a future it can feel but never hold.

I do not know if that is true. I know I keep pulling at it.

Key takeaways

Observation is any interaction that leaves a record. If anything could tell which path the light took, the wave pattern fades in proportion, whether or not anyone looks.
The cut is movable. Physicists call it the Heisenberg cut, and the line between the measured and the measurer can be slid without changing the predictions.
A water wave is a crowd behavior. It exists only across many interacting parts, which is why films simulate water with millions of particles. Light is the exception: its quantum wave belongs to each photon.
Language models brace, they do not fetch. Weights are frozen at run time; the hidden state arranges itself around the expected next word.
Memory works like a stored forecast. Predict, observe, fold the gap into the next prediction: the Kalman filter, the remembering brain, and a good agent handoff all run that loop.

Sources

Lamb, W. E., Jr., & Scully, M. O. (1969). The photoelectric effect without photons. In Polarisation, Matière et Rayonnement. Presses Universitaires de France. https://ntrs.nasa.gov/citations/19690054849
Grangier, P., Roger, G., & Aspect, A. (1986). Experimental evidence for a photon anticorrelation effect on a beam splitter. Europhysics Letters 1(4): 173-179. https://doi.org/10.1209/0295-5075/1/4/004
Dürr, S., Nonn, T., & Rempe, G. (1998). Origin of quantum-mechanical complementarity probed by a which-way experiment in an atom interferometer. Nature 395: 33-37. https://doi.org/10.1038/25653
Hornberger, K., Uttenthaler, S., Brezger, B., Hackermüller, L., Arndt, M., & Zeilinger, A. (2003). Collisional decoherence observed in matter wave interferometry. Physical Review Letters 90: 160401. https://doi.org/10.1103/PhysRevLett.90.160401
Englert, B.-G. (1996). Fringe visibility and which-way information: An inequality. Physical Review Letters 77: 2154-2157. https://doi.org/10.1103/PhysRevLett.77.2154
Bacciagaluppi, G. (2003, rev. 2025). The role of decoherence in quantum mechanics. Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/qm-decoherence/
Faye, J. (rev. 2024). Copenhagen interpretation of quantum mechanics. Stanford Encyclopedia of Philosophy (on the Heisenberg cut and von Neumann, 1932). https://plato.stanford.edu/entries/qm-copenhagen/
Piazza, L., Lummen, T. T. A., Quiñonez, E., Murooka, Y., Reed, B. W., Barwick, B., & Carbone, F. (2015). Simultaneous observation of the quantization and the interference pattern of a plasmonic near-field. Nature Communications 6: 6407. https://doi.org/10.1038/ncomms7407 For why the particle half needs more than this experiment: Dahan, R., et al. (2021). Imprinting the quantum statistics of photons on free electrons. Science 373. https://doi.org/10.1126/science.abj7128
Müller, M., Charypar, D., & Gross, M. (2003). Particle-based fluid simulation for interactive applications. Proc. ACM SIGGRAPH/Eurographics Symposium on Computer Animation: 154-159. https://doi.org/10.2312/SCA03/154-159
Van Opstal, B., Janin, L., Museth, K., & Aldén, M. (2014). Large scale simulation and surfacing of water and ice effects in Dragons 2. ACM SIGGRAPH 2014 Talks. https://doi.org/10.1145/2614106.2614156
Reynolds, C. W. (1987). Flocks, herds and schools: A distributed behavioral model. ACM SIGGRAPH Computer Graphics 21(4): 25-34. https://doi.org/10.1145/37402.37406
Littman, M. L., Sutton, R. S., & Singh, S. (2001). Predictive representations of state. NeurIPS 14. https://papers.nips.cc/paper/2001/hash/1e4d36177d71bbb3558e43af9577d70e-Abstract.html
Stachenfeld, K. L., Botvinick, M. M., & Gershman, S. J. (2017). The hippocampus as a predictive map. Nature Neuroscience 20: 1643-1653. https://doi.org/10.1038/nn.4650
Perdomo, J. C., Zrnic, T., Mendler-Dünner, C., & Hardt, M. (2020). Performative prediction. ICML 2020. https://arxiv.org/abs/2002.06673 See also Merton, R. K. (1948). The self-fulfilling prophecy. The Antioch Review 8(2): 193-210. https://doi.org/10.2307/4609267
Brown, T. B., et al. (2020). Language models are few-shot learners. NeurIPS 2020. https://arxiv.org/abs/2005.14165
Lindsey, J., Gurnee, W., Ameisen, E., et al. (2025). On the biology of a large language model. Transformer Circuits Thread. https://transformer-circuits.pub/2025/attribution-graphs/biology.html
Wu, W., Morris, J. X., & Levine, L. (2024). Do language models plan ahead for future tokens? COLM 2024. https://arxiv.org/abs/2404.00859
Ba, J., Hinton, G., Mnih, V., Leibo, J. Z., & Ionescu, C. (2016). Using fast weights to attend to the recent past. NeurIPS 2016. https://arxiv.org/abs/1610.06258
Behrouz, A., Zhong, P., & Mirrokni, V. (2025). Titans: Learning to memorize at test time. https://arxiv.org/abs/2501.00663
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic Engineering 82: 35-45. https://doi.org/10.1115/1.3662552
Schacter, D. L., Addis, D. R., & Buckner, R. L. (2007). Remembering the past to imagine the future: The prospective brain. Nature Reviews Neuroscience 8: 657-661. https://doi.org/10.1038/nrn2213
Hassabis, D., Kumaran, D., Vann, S. D., & Maguire, E. A. (2007). Patients with hippocampal amnesia cannot imagine new experiences. PNAS 104(5): 1726-1731. https://doi.org/10.1073/pnas.0610561104
Sevenster, D., Beckers, T., & Kindt, M. (2013). Prediction error governs pharmacologically induced amnesia for learned fear. Science 339: 830-833. https://doi.org/10.1126/science.1231357
Addis, D. R., Wong, A. T., & Schacter, D. L. (2007). Remembering the past and imagining the future: Common and distinct neural substrates during event construction and elaboration. Neuropsychologia 45: 1363-1377. https://doi.org/10.1016/j.neuropsychologia.2006.10.016
Cover, T. M., & Thomas, J. A. (2006). Elements of Information Theory (2nd ed.), Theorem 2.8.1, the data processing inequality. Wiley. https://doi.org/10.1002/047174882X
Hensen, B., et al. (2015). Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres. Nature 526: 682-686. https://doi.org/10.1038/nature15759
Ghirardi, G. C., Rimini, A., & Weber, T. (1980). A general argument against superluminal transmission through the quantum mechanical measurement process. Lettere al Nuovo Cimento 27: 293-298. https://doi.org/10.1007/BF02817189
Rovelli, C. (1996). Relational quantum mechanics. International Journal of Theoretical Physics 35: 1637-1678. https://doi.org/10.1007/BF02302261

The quantum language here (the cut, the observer, memory as a collapsed forecast) is metaphor and intuition pump, not a physics claim. The sources mark where the essay touches real results, and the no-communication theorem [27] is named precisely because it rules out the wildest version. The agent board in “Memory is a prediction” is my own working practice.

Related reading: There Is a New Voter in the Room is the other half of this. That one points the cut at words: a finished sentence voted into fact by people, and a language model that has to live with every word it writes. This one points it at perception, a single mind meeting the world. The Orphan That Learned to Say I pushes on the same nerve from the inside: what happens when the predicting system turns and models the predictor. And Saving Is Older Than Planning is the honest counterweight from my own small universe: creatures that never learned to see winter coming still learned to save for it, no forecast required. ZCoP: The Plan That Rewrites Itself gives the next field a whole table: every finished step collapsed to facts and the next action written down, so a restarted agent resumes from the plan instead of the transcript.

The Loop of Attention is the same question inside one session: how an agent loses its goal long before any restart, and how reciting the goal puts it back.