The Extended Mind

Executive Summary

In 1998, two philosophers published a nine-page paper in the journal *Analysis* that argued something most people found absurd: your mind is not entirely inside your head.

Andy Clark and David Chalmers made the case with a thought experiment. Inga hears about an exhibition at the Museum of Modern Art. She recalls that the museum is on 53rd Street, and she walks there. Otto has Alzheimer's disease. He carries a notebook everywhere and writes new information into it. When Otto hears about the exhibition, he consults the notebook, finds the entry for the museum's address, and walks to 53rd Street.

Clark and Chalmers asked: what is the principled difference? Inga's belief was in biological memory. Otto's belief was in a notebook. Both were reliably available, both were consulted automatically, both drove the same behavior. If we call Inga's notebook entry a belief because it sits in neurons, and refuse to call Otto's a belief because it sits in ink, we are drawing the line at the skull for no reason other than that the skull is where we are used to drawing it.

That paper launched what is now called the extended mind thesis. Nearly three decades later, it has stopped being a philosophy seminar question and become an infrastructure question.

Because in 2026, almost everyone is Otto. The notebook is a cloud service. The entries are conversations, documents, and inferred facts about your work. The retrieval is a similarity search you cannot inspect. And the notebook is owned, priced, versioned, and deprecated by a company that has never met you.

If Clark and Chalmers were right, then the terms of service on your AI assistant are not a commercial document. They are a set of conditions under which part of your cognition is permitted to keep operating.

1. The Argument

1.1 The Parity Principle

The extended mind thesis rests on a single test that Clark and Chalmers called the parity principle.

If, as we confront some task, a part of the world functions as a process which, were it to go on in the head, we would have no hesitation in recognizing as part of the cognitive process, then that part of the world is part of the cognitive process.

The principle is deliberately austere. It does not ask whether the process is biological. It does not ask whether it is conscious. It asks only whether we would count it as cognition if it happened internally. If the answer is yes, then location is not a reason to change the verdict.

This is a claim about where the boundary of a cognitive system falls, not a claim about consciousness. Clark and Chalmers were careful about this. Otto's notebook is part of his memory. It is not part of his experience. Nobody argues that the paper feels anything.

1.2 Active Externalism

Philosophy already had an externalism before 1998. Hilary Putnam and Tyler Burge had argued that the *content* of mental states depends on facts outside the head: what your word "water" means depends on the chemistry of the stuff around you, not on anything in your skull. But that externalism was passive. The environment fixed the meaning and then sat still.

Clark and Chalmers proposed something more aggressive, which they called active externalism. The environment does not merely fix content from a distance. It participates, in real time, in the processing. The relevant unit of analysis is not the brain but the coupled system: brain plus tool, operating together, where removing the tool degrades the system's behavior the way removing a chunk of cortex would.

The test for coupling is not sentiment. It is what happens on removal. Take Inga's hippocampus and she cannot find the museum. Take Otto's notebook and he cannot find the museum. The functional consequence is identical, which is exactly the point.

1.3 The Four Conditions

Clark and Chalmers did not claim that every external object joins your mind. A billboard you glance at once is not part of your cognition. They set out conditions the resource has to meet, sometimes called the glue-and-trust criteria:

1. Constant availability. The resource is a constant in your life. When the relevant information is needed, you rarely act without consulting it.
2. Easy accessibility. The information is directly available without difficulty. There is no significant friction between wanting it and having it.
3. Automatic endorsement. When you retrieve information from the resource, you endorse it automatically. You do not treat it as a claim to be evaluated. You treat it as something you know.
4. Prior conscious endorsement. The information is there because you put it there, or approved it, at some earlier point.

Clark and Chalmers flagged the fourth condition as the most contentious, and it is the one that has aged the strangest. In 1998, condition four was nearly automatic: nothing entered Otto's notebook unless Otto wrote it. In 2026, systems write to your external memory constantly, without your knowledge, from sources you never reviewed.

Hold onto those four conditions. Everything that follows is an argument about what happens when software satisfies some of them and quietly fails the others.

2. The Evidence

The extended mind thesis is a philosophical claim, but the behavior it describes is empirically well documented. People really do treat external stores as memory, and they reorganize their internal cognition around that assumption.

2.1 The Google Effect

Sparrow, Liu, and Wegner (2011) ran four experiments on how people remember information they expect to be able to look up. Participants who were told that the statements they typed would be saved recalled fewer of the statements themselves and were significantly better at recalling *where* the statements had been stored. Participants who were told the file would be erased remembered the content.

The finding is not that search engines make people stupid. It is that memory reallocates. When an external store is reliable, the brain stops encoding content and starts encoding addresses. This is exactly the behavior the extended mind thesis predicts: the coupled system optimizes across both halves, and the internal half specializes in indexing.

2.2 Saving Makes You Smarter On the Next Thing

Storm and Stone (2015) gave participants two files of study material. Half were allowed to save the first file before studying the second. Those who saved recalled substantially more of the *second* file than those who did not.

Offloading was not a trade. It was a net gain. The cognitive resources tied up in maintaining the first file were released the moment the participants trusted that the file was safe. Critically, the benefit disappeared when participants had reason to doubt that saving had worked. Unreliable external memory is worse than no external memory, because you pay the maintenance cost anyway and you do not get to redeploy the freed capacity.

2.3 The World as Its Own Best Model

Ballard, Hayhoe, Pook, and Rao (1997) tracked eye movements while people copied a pattern of colored blocks. The efficient strategy, if you were designing a robot, would be to memorize the model and then build. That is not what people did. They looked at the model, picked up one block, looked at the model again, placed the block, looked again. They re-fixated the same block sometimes twice: once for color, once for position.

Human beings do not memorize when looking is cheaper. They use the world as external storage with a latency of roughly 200 milliseconds, and they let their eyes do the retrieval. Ballard called this a deictic strategy. It is the extended mind in action, measured in saccades.

2.4 Epistemic Action

Kirsh and Maglio (1994) studied expert Tetris players and noticed they rotated pieces more often than the game required. The extra rotations were not mistakes. Physically rotating a piece took roughly 300 milliseconds. Mentally rotating the same piece took roughly 1,000. Players were offloading the rotation to the game because the game was faster than their imagination.

Kirsh and Maglio named this *epistemic action*: a physical action performed not to advance the task but to make cognition easier. The keyboard shortcut that saves you a mental step is an epistemic action. So is highlighting a paragraph before you think about it.

This matters enormously for design, and section 5 returns to it: whether people extend their minds into a tool is decided by a latency comparison. If consulting the tool is slower than thinking, they will think. The tool never becomes part of the system.

2.5 Transactive Memory

Wegner (1987) documented that couples and long-running teams distribute memory across people. Each member becomes the designated store for a domain, and the group's total recall exceeds what any member holds. Members do not remember the facts. They remember who remembers the facts.

Wegner also documented the failure mode. When a transactive memory system is broken by separation or departure, the remaining members do not simply lose access to information. They experience something closer to loss of self, because a chunk of their functional memory has walked out the door.

Every organization that has lost the one person who knew how the billing system worked has run this experiment.

3. Your AI Is Otto's Notebook

Apply the four conditions to a modern cloud AI assistant with persistent memory, and the result is uncomfortable.

Conditions two and three are satisfied by essentially every AI assistant on the market. That is what makes them feel like thinking rather than like using software. But conditions one and four are where cloud architecture and cognitive extension come apart.

3.1 The Availability Failure

Otto's notebook has never had an outage. It has never been deprecated in favor of Notebook 4 Turbo. Its terms of service have never changed. No government has ever restricted its export.

The relevant question is not whether cloud AI is usually available. It usually is. The question is what class of events can take it away, and whether those events are within your control. The current list includes: service outage, model deprecation, price change beyond what you will pay, account suspension for a policy determination you cannot appeal, corporate acquisition, corporate failure, regional restriction, and export control. None of them are within your control. Several of them arrive with no notice.

Our companion papers *The Day the Model Disappeared* and *Open Source as Insurance* cover the 2026 disruptions in detail. The point here is narrower and more structural: an external resource that can be revoked by a third party does not satisfy Clark and Chalmers' first condition. It is not glued. It is rented.

And because condition one is what licenses conditions two and three, revocability contaminates the whole arrangement. You have organized your work around a store you do not control, and the more successfully you have extended into it, the more of your functional memory leaves with it.

3.2 The Endorsement Failure

Condition four is the one nobody notices, and it is the one that has genuinely inverted since 1998.

Otto's notebook contained only what Otto wrote. Every entry passed through his judgment before it was recorded. When he retrieved an entry and endorsed it automatically, that automatic endorsement was earned: he was trusting a past version of himself who had, at the time, checked.

Modern AI memory does not work this way. Systems extract entities from your conversations, infer preferences from your behavior, summarize documents you skimmed, and write all of it into a store you never review. Then, on retrieval, condition three fires exactly as Clark and Chalmers described. You endorse it automatically. It arrives with the phenomenology of remembering, not the phenomenology of being told.

This is a genuinely new failure mode. The mechanism of extension is working correctly. The trust that the mechanism assumes has been quietly removed. Our paper *The Invisible Anchor* traces what happens downstream: retrieved facts function as anchors, similarity scores get read as confidence, and asking the system to double-check its own output reinforces the anchor rather than testing it.

4. Whose Mind Is It

If the thesis is right, the practical consequences are not philosophical. They are legal, commercial, and ethical.

4.1 Ethical Parity

Neil Levy (2007) proposed the ethical parity principle as the moral counterpart to Clark and Chalmers' cognitive one: if external and internal components of a cognitive system are functionally on a par, then interventions on them are ethically on a par. Altering someone's external memory is not categorically different from altering their internal memory. It is the same act, performed on a different substrate.

Carter and Palermos (2016) took the argument to its conclusion in the *Journal of the American Philosophical Association*, asking whether compromising someone's computer or phone constitutes personal assault rather than property damage. If a device is genuinely constitutive of a person's cognition, then breaking into it is not analogous to breaking into their car. It is analogous to breaking into them.

Clark and Chalmers themselves gestured at this in the paper's closing pages, noting that on their view, interfering with someone's environmental resources can carry the moral weight of interfering with the person.

4.2 What This Reframes

Take that seriously for five minutes and ordinary product decisions stop looking ordinary.

Vendor access to your memory store is not a data-processing question. It is third-party read access to a part of your cognition.

Silent model swaps are not a version-management question. The retrieval and reasoning behavior of your external memory changed without your consent, which is closer to a neurological event than a release note.

Memory deletion on cancellation is not a data-retention question. Our paper *The Memory Trap* covers the commercial dynamics; the extended-mind framing sharpens it. A subscription that holds your accumulated knowledge hostage is not selling you a tool. It is renting you access to something you built and that now does part of your thinking.

Training on your stored context is not a licensing question. It is the extraction of a cognitive component's contents for a purpose you did not authorize.

None of this requires the metaphysics to be settled. It requires only that the coupling be real, which section 2 established empirically.

4.3 The Honest Counterarguments

The extended mind thesis is contested, and a paper that pretended otherwise would be doing exactly the automatic endorsement it warns about.

The coupling-constitution fallacy. Adams and Aizawa (2001, 2008) argue that Clark and Chalmers confuse a causal relationship with a constitutive one. Your cognition is causally coupled to your notebook the way a fire is causally coupled to oxygen, but nobody says the atmosphere is part of the fire. They also propose a mark of the cognitive that external resources fail: genuine cognitive states carry intrinsic, non-derived content, whereas the marks in a notebook mean something only because someone interprets them.

The explanatory redundancy objection. Rupert (2004) argues that everything the extended mind thesis explains is explained equally well by embedded cognition: the claim that cognition depends heavily on external scaffolding, without the claim that the scaffolding is part of the mind. He notes that external memory behaves unlike internal memory in measurable ways. It shows no generation effect and no interference patterns. If the two are so functionally different, why file them under one category?

The dependency objection. A separate line of criticism holds that extension is a polite word for atrophy. Our paper *Cognitive Surrender* takes this seriously.

The response to all three is the same, and it is a design response rather than a philosophical one. Whether the notebook is literally constitutive of Otto's mind or merely something he is deeply dependent on, the failure modes are identical. Losing it degrades him. Corrupting it misinforms him. Reading it exposes him. You do not need to win the metaphysics to conclude that a resource this deeply integrated should be reliable, private, auditable, and owned by the person using it.

The distinction that survives all three objections is between extension and surrender. Extension keeps the endorsement with the person: the system holds what you decided, and you can see why. Surrender moves the endorsement into the system: it decides, and you accept. The four conditions are what separates them, and condition four is the load-bearing one.

5. Designing for Extension

Heersmink (2015) argued that integration between a person and a cognitive artifact is not binary but graded along measurable dimensions: information flow, reliability, durability, trust, procedural transparency, informational transparency, individualization, and transformation. That framework turns "is this part of my mind" into an engineering checklist.

Here is that checklist applied to an AI system, and what each item actually requires.

5.1 Reliability and Durability

Condition one is availability, and availability is architecture. A system that cannot run without a network, a subscription, and a vendor's continued goodwill fails the condition no matter how good it is on the days it works.

Kent stores its knowledge graph as a local database file on the user's own disk, encrypted at rest, alongside the archived content and the connector tokens. Private mode runs inference locally through Ollama and makes zero outbound network requests. Cloud providers are one option among several rather than a dependency, and the same skills run against any of six providers, so a provider disappearing changes cost and quality but not capability.

Durability also means the store outlives the vendor. A memory you cannot export is a memory with a termination date attached.

5.2 Accessibility, Which Is Really Latency

Kirsh and Maglio's Tetris result is the most underrated finding in this literature for anyone building software. Whether a tool becomes part of your cognition is decided by a latency comparison your brain runs without telling you. If consulting the tool costs more than thinking unaided, you will think unaided, and the tool will never be more than an application you occasionally open.

This is why the interface question is a philosophical question and not a cosmetic one. Open a browser, find the tab, wait for load, paste the context, restate what you are doing: that is thirty seconds of overhead, and thirty seconds is far past the threshold where offloading stops being worth it. Our paper *The Cost of the Tab* quantifies the switching cost directly.

Kent's answer is the overlay: highlight text anywhere on the system, press Ctrl+Shift+Space, act. There is no destination. The cost of consulting the external store drops close enough to the cost of a thought that the comparison stops mattering, and the coupling actually forms.

Time to consult an external store, in seconds:

5.3 Trust, Earned Rather Than Assumed

Condition three says you will endorse retrieved information automatically. You cannot design that away; it is what makes extension feel like memory. What you can do is make the automatic endorsement deserved.

Kent does this in four places. Every node in the graph carries a confidence score that decays exponentially with time since last confirmation, so a fact from fourteen months ago does not arrive looking identical to one from yesterday. Memory is tiered, and old unreferenced knowledge loses embedding precision and eventually loses semantic retrieval entirely, so stale facts surface less rather than more. The retrieval instruction frames memory as hypothesis rather than ground truth. And for high-stakes queries, Deep Mode routes to a second model from a different provider that does not share the first model's context, so agreement means something.

5.4 Informational Transparency, or Condition Four Restored

The 1998 version of condition four was trivially satisfied because Otto did the writing. Restoring it in software means making the provenance of every stored fact recoverable.

Kent writes informed_by edges from each response back to the specific nodes that informed it. That produces a traceable chain: this answer came from that fact, which came from that email, on that date. When a fact turns out to be wrong, you do not just correct the fact. You walk the edges and find every conclusion that depended on it.

Kent also applies a write discipline gate: it does not store what it can re-derive from a connected source. Meeting times live in your calendar and are fetched live. Email subjects live in Gmail. What gets written into the graph is what cannot be re-derived, which keeps the store small enough that provenance stays inspectable.

5.5 Individualization and Privacy

Heersmink's individualization dimension asks whether the artifact is shaped to the specific person. A general-purpose model that treats you as a member of a population is a tool. A store that has accumulated your projects, your terminology, your commitments, and your working style is closer to a component.

That is precisely why Levy's ethical parity principle bites hardest here. The more individualized the store, the more reading it resembles reading a mind rather than reading a database. Our paper *Thinking in Private* makes the case that some thinking only happens when nobody is watching; the extended mind thesis explains why that is not paranoia. If the store is part of the cognition, then surveillance of the store is surveillance of the thinking.

Kent's private mode exists for this reason: local model, local graph, zero outbound requests, with full connector parity so that choosing privacy does not mean choosing a weaker assistant.

6. What Changes If You Take This Seriously

Three practical consequences follow, and none of them require agreeing with Clark and Chalmers about metaphysics.

Evaluate AI tools on revocability, not just capability. Capability comparisons are what vendors compete on and what benchmarks measure. Revocability is what determines whether the thing you build on top of the tool survives. Ask what happens to your accumulated context if the vendor changes price, changes policy, changes hands, or goes away. If the answer is that you lose it, you are not extending your mind. You are leasing a prosthetic.

Treat unreviewed writes as the real risk. The industry worries about hallucination, which is visible and catchable. The extended mind framing points at something quieter: facts that enter your external memory without your review and are then retrieved with the full authority of things you know. Demand provenance. A memory system that cannot tell you where a fact came from has failed condition four, and condition four is what separates memory from suggestion.

Measure the friction, because friction decides everything. If your AI takes thirty seconds to consult, you will not consult it for the small questions, and the small questions are most of thinking. A tool that is brilliant and slow loses to a tool that is good and instant, not because users are impatient but because the brain is running a latency comparison it will not let you override.

Conclusion

Clark and Chalmers wrote their paper in an era when the strongest example they could reach for was a man with a notebook. They picked Otto deliberately, because Alzheimer's made the dependency vivid. The point was that the same structure was present, less dramatically, in everyone.

They were right about that, and the intervening decades have made the example look quaint. The notebook is now a service. It writes to itself. It retrieves by a mechanism you cannot inspect. It is priced monthly and can be discontinued. And it is used by people whose biological memory is fine, which was always the more interesting case.

Otto's notebook satisfied every condition Clark and Chalmers set. It was always available, instantly accessible, automatically trusted, and everything in it was there because Otto put it there. It is a strange thing that the 1998 example clears a bar that most 2026 systems do not.

The gap is not a technical limitation. Every condition is achievable. Availability is achieved by running locally and owning the store. Accessibility is achieved by removing the destination. Earned endorsement is achieved by decay, skeptical retrieval, and independent verification. Conscious endorsement is achieved by making provenance traceable and refusing to store what does not need storing.

These are architecture decisions, and they cost something, which is why most systems do not make them. They are invisible when they work. Nobody notices the outage that could not happen or the anchor that decayed before it did damage.

But if the thesis holds, and thirty years of evidence suggests something like it does, then these are not product preferences. Every decision about where your knowledge lives, who can read it, whether it can be revoked, and whether you can see where it came from is a decision about the boundary of your own mind, made by someone.

The only real question is whether that someone is you.

References

1. Clark, A. & Chalmers, D. (1998). "The Extended Mind." *Analysis*, 58(1), 7-19.

1. Clark, A. (2008). *Supersizing the Mind: Embodiment, Action, and Cognitive Extension*. Oxford University Press.

1. Menary, R. (Ed.) (2010). *The Extended Mind*. MIT Press.

1. Sparrow, B., Liu, J., & Wegner, D.M. (2011). "Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips." *Science*, 333(6043), 776-778.

1. Storm, B.C. & Stone, S.M. (2015). "Saving-Enhanced Memory: The Benefits of Saving on the Learning and Remembering of New Information." *Psychological Science*, 26(2), 182-188.

1. Risko, E.F. & Gilbert, S.J. (2016). "Cognitive Offloading." *Trends in Cognitive Sciences*, 20(9), 676-688.

1. Gilbert, S.J. (2015). "Strategic Offloading of Delayed Intentions into the External Environment." *Quarterly Journal of Experimental Psychology*, 68(5), 971-992.

1. Ballard, D.H., Hayhoe, M.M., Pook, P.K., & Rao, R.P.N. (1997). "Deictic Codes for the Embodiment of Cognition." *Behavioral and Brain Sciences*, 20(4), 723-767.

1. Kirsh, D. & Maglio, P. (1994). "On Distinguishing Epistemic from Pragmatic Action." *Cognitive Science*, 18(4), 513-549.

1. Wegner, D.M. (1987). "Transactive Memory: A Contemporary Analysis of the Group Mind." In B. Mullen & G.R. Goethals (Eds.), *Theories of Group Behavior*, 185-208. Springer-Verlag.

1. Adams, F. & Aizawa, K. (2001). "The Bounds of Cognition." *Philosophical Psychology*, 14(1), 43-64.

1. Adams, F. & Aizawa, K. (2008). *The Bounds of Cognition*. Blackwell.

1. Rupert, R.D. (2004). "Challenges to the Hypothesis of Extended Cognition." *The Journal of Philosophy*, 101(8), 389-428.

1. Levy, N. (2007). *Neuroethics: Challenges for the 21st Century*. Cambridge University Press.

1. Carter, J.A. & Palermos, S.O. (2016). "Is Having Your Computer Compromised a Personal Assault? The Ethics of Extended Cognition." *Journal of the American Philosophical Association*, 2(4), 542-560.

1. Heersmink, R. (2015). "Dimensions of Integration in Embedded and Extended Cognitive Systems." *Phenomenology and the Cognitive Sciences*, 14(3), 577-598.

Published by Kent Research, July 2026. This paper draws on published work in philosophy of mind and cognitive psychology. Descriptions of Kent's architecture reflect the shipping product. Nothing here constitutes professional advice.