What counts as genuine memory under the Extended Mind thesis?
This explores what the Extended Mind thesis would count as 'real' memory — whether storage outside the skull (a notebook, a retrieval database, even the layout of a room) qualifies as genuine remembering, and how that bar applies to AI systems.
This explores what the Extended Mind thesis would actually count as genuine memory — and the corpus is interesting precisely because it shows the thesis being applied, contested, and quietly walked back. The original 1998 Clark and Chalmers argument says memory isn't defined by where it sits (inside a brain) but by the functional role it plays: if an external resource is reliably available, automatically trusted, and easily accessed when you need it, it counts as remembering just as much as a biological recall does. The notebook in your pocket can be part of your mind. What's striking is that one of the co-authors appears to have reversed course: the Did Chalmers abandon his own Extended Mind principles? note argues that Chalmers' 2026 account of LLMs quietly relocates the thinking 'inside' the AI system, adopting exactly the internalist, skull-bound boundary the 1998 thesis was built to reject. So the first thing the reader learns is that 'genuine memory' under this thesis was never about a place — and that even its authors find that hard to hold onto.
If you take the functional criterion seriously, some surprising things start to count. The Do RL agents accidentally use environments as memory? note offers a clean, almost startling case: reinforcement-learning agents that leave traces in their environment — marks, paths, arrangements of objects — end up *reading those traces back* to reconstruct their own history, and a mathematical proof shows this genuinely reduces the information they'd otherwise have to carry internally. Nobody programmed them to 'remember' this way; it fell out of ordinary reward optimization. That's the Extended Mind thesis in miniature: the world itself becomes the storage medium, and the boundary between 'agent' and 'memory' dissolves. It satisfies the situated-cognition criteria without any explicit memory module at all.
The contrast with how AI systems are usually built is sharp. The Can brain memory systems explain how LLMs should store knowledge? note maps an internalist picture instead: transformer weights as a slow 'neocortex' of consolidated knowledge, a RAG retrieval store as a fast 'hippocampus,' and agentic working state as 'prefrontal' executive control. This is a useful engineering frame, but notice it keeps all the memory *inside the system* and merely tiers it — the opposite move from the extended-mind view, where the environment is a legitimate tier in its own right. Reading these two notes against each other surfaces the real question: is genuine memory a property of stored representations, or of a reliable coupling between an agent and resources it can count on?
The corpus also hints at why the boundary question matters beyond bookkeeping. The Can disembodied language models ever qualify as conscious? note argues that mental language — including, plausibly, 'remembering' as opposed to mere 'retrieving' — originates from embodied creatures who share a world and triangulate on common objects. On that view, what makes memory *genuine* rather than mechanical lookup isn't the storage at all but the situated, world-coupled relationship to what's stored — which is exactly the ingredient a disembodied database lacks and an environment-coupled RL agent partially has. And Can we defend modest mental attributions to large language models? suggests a graded answer rather than a yes/no one: we might ascribe undemanding mental states to systems without granting them rich inner experience, the way we already do for animals.
The thing you may not have known you wanted to know: under a strict reading of the Extended Mind thesis, a retrieval-augmented language model — the architecture everyone calls 'giving the AI a memory' — might *not* count as genuinely remembering, while a humble RL agent scuffing marks into its environment might. Genuineness lives in the coupling, not the cabinet.
Sources 5 notes
The 2026 virtual-instance account locates the LLM interlocutor inside the AI system, implicitly adopting internalist boundaries that the 1998 Extended Mind thesis explicitly rejected. This creates internal inconsistency unless the earlier thesis is retracted or the new application misapplies its principles.
Mathematical proof shows that environmental artifacts reduce information needed to represent history in RL agents. Path-following agents naturally develop memory-like behavior through standard reward optimization, satisfying situated cognition criteria without explicit memory objectives.
Research shows transformer weights function as a distributed neocortex for consolidated knowledge, RAG stores as hippocampal indexing for rapid encoding, and agentic state as prefrontal executive control. The CLS framework predicts why hybrid systems outperform single-tier approaches and identifies missing consolidation mechanisms that prevent memory integration.
Current disembodied LLMs cannot be candidates for consciousness because consciousness language originates from and applies only to entities sharing a world with us through co-presence and triangulation on shared objects.
Both robustness and etiological deflationist arguments beg the question against inflationism. A graded approach ascribing metaphysically undemanding states like beliefs and desires—while withholding consciousness claims—mirrors how we treat non-human animals.