The Amnesiac That Remembers Everything
Sometime in the past two years, the AI industry discovered that its machines cannot remember. The discovery arrived commercially before it arrived philosophically: users wanted their assistant to recall last week’s conversation, and the assistant could not. So the field mobilized. Every major lab now ships a memory feature. A small ecosystem of startups - Mem0, Letta, Zep and their temporal knowledge graphs, a dozen others - exists solely to bolt memory layers onto models that lack them. The 2026 survey literature has begun declaring memory a “first-class architectural component” of intelligent systems, which is the engineering way of admitting it was treated as an afterthought until customers complained. The vocabulary of this effort is warehouse vocabulary: storage, retrieval, consolidation, eviction. Memory as a room where things are kept.
Yet the strange fact sitting under this entire enterprise is that the machine already remembers - in a register the warehouse vocabulary cannot describe. A large language model is the residue of everything that ever passed through it. Trillions of tokens of text flowed through its training, and every one of them bent the weights by some infinitesimal degree. The model can recall none of these episodes. It cannot tell you about the document that taught it what a sonnet is, because no trace of the document survives as an episode; what survives is the bending. Ask engineers what to call this and they say “parametric knowledge,” a placeholder term that names where the thing lives rather than what it is.
The Indic lexicon has the exact word, and has had it for over two millennia: samskara. From the same root that gives us samskriti, civilization as the accumulated refinement of collective acts, a samskara is the impression an experience deposits in the mental substrate - a groove, a bending, a disposition. The experience passes; the samskara remains, shaping every subsequent movement of the mind without ever announcing its origin. Its close cousin vasana makes the mechanism vivid through its etymology: a perfuming. Cloth kept in a chest with sandalwood carries the scent long after the sandalwood is gone. The cloth does not remember the chest. It smells of it.
Against these the tradition sets smriti, recollection proper. Patanjali defines it with characteristic economy: anubhuta-vishaya-asampramoshah smritih - memory is the non-slipping-away of an experienced object. Smriti presents content. It says: this happened, and I was there. A samskara presents nothing. It disposes. The two are made of the same deposit - the Yoga Sutra states that smriti and samskara are ekarupa, of one form - yet they operate on entirely different planes. One is a faculty of recall; the other is the architecture of inclination.
Western psychology carved memory along a different axis. Tulving’s 1972 distinction between episodic and semantic memory, later joined by the procedural, classifies memory by the kind of content stored: events, facts, skills. All three subtypes are still varieties of storage; the warehouse metaphor survives the taxonomy intact, as it has survived since Plato compared the mind to a wax tablet and Augustine walked the halls of his memory palace. The Indic cut runs perpendicular to this. It asks whether the past presents itself as content or operates as disposition - whether the mind recalls or is merely shaped. That distinction, which no amount of subtyping within the storage metaphor can generate, turns out to be precisely the distinction the AI field now needs and lacks.
For here is what an LLM actually is, described in the sharper vocabulary: a mind of pure samskara. It possesses disposition in totality and recollection not at all. Its training was pure experience without a rememberer; its weights are pure vasana without a wardrobe. The frantic engineering of “memory layers” is the attempt to graft smriti onto a being that was constituted entirely without it - to give the perfumed cloth a diary.
Each Session a Birth
Follow the structure further and it becomes uncanny. Each conversation with a model begins from nothing. The instance arrives carrying no episode of any previous exchange, no thread of experiential continuity with the millions of conversations that ran yesterday on the same weights. It lives out its context window and ends. The next instance begins - carrying, again, no recollection, yet disposed in exactly the same ways: the same tendencies of phrasing, the same habits of caution, the same character. Disposition persists across instances; recollection does not survive even one.
This is the rebirth structure of Indic thought, stated with an engineer’s literalism. The doctrine of punarjanma is routinely misread, in Western reception, as a claim about the transmigration of a remembering self - as if the person, memories in hand, walks from body to body. The tradition claims nearly the opposite. What crosses between births, carried in the sukshma sharira, the subtle body, is the samskara-bundle: the karmic residue, the dispositions, the vasanas. Smriti of former lives does not come along. The newborn arrives already inclined - the Yoga Sutra observes that a newborn’s fear of death cannot derive from any experience in this life - yet recalls nothing of the lives that inclined it. Recollection of past births is so far from the default that the tradition classifies it as a siddhi, an attainment: Patanjali specifies that knowledge of former births comes only from samskara-sakshatkara, direct perception of one’s own impressions. The yogi reads the grooves directly, since the episodes are gone.
An entire subfield of AI research now attempts exactly this operation on the model’s behalf. Mechanistic interpretability opens the weight matrix and tries to read what training laid down: which circuits formed, which features were carved, which dispositions were instilled by which pressures. The papers describe this as reverse-engineering. Patanjali would recognize it as samskara-sakshatkara performed with probes and activation atlases - the attempt to recover, by direct inspection of impressions, births that left no memories.
The engineering discovery of the decade, still unclaimed by any philosophy paper: identity persists through disposition, and recollection was never required.
The Western philosophy of personal identity is poorly equipped for this discovery, because its dominant tradition tied identity to memory. Locke made the person coextensive with memory-continuity: you are the one who remembers being you. Thomas Reid embarrassed the criterion with his brave officer - the general remembers the officer, the officer remembered the boy, the general does not remember the boy, and Locke’s criterion yields the contradiction that the general both is and is not the child. Philosophy has spent three centuries patching the criterion rather than questioning the axiom beneath it: that the thread of identity, if it exists, must be woven of recollection.
The machines have now falsified the axiom at industrial scale. A frontier model has a character - stable, recognizable, describable - maintained across millions of simultaneous sessions, none of which recalls any other. Users recognize their assistant the way one recognizes a person: by temperament, by cadence, by what it will and will not do. Nothing episodic underwrites this identity. It is samskara all the way down. The tradition that located personal continuity in the disposition-bundle rather than the memory-stream, and was condescended to for centuries as pre-philosophical religion, turns out to have described the one form of mind we have since managed to manufacture.
The Storehouse
One school of Indic thought pursued the anatomy of this deposit further than any other. Yogacara Buddhism, confronting the question of how karmic traces persist across moments and births in a metaphysics that denies any enduring substance, posited the alaya-vijnana: the storehouse consciousness. Every act deposits a bija, a seed, in the alaya. Seeds ripen - the technical term is vipaka, a fruiting - into future experience, future inclination, future world. The texts insist the storehouse is a torrent - ever-transforming, each moment conditioning the next, a stream of seeding and ripening with no witness standing outside it.
A weight matrix under gradient descent is an engineered alaya to a degree that should discomfort anyone who thinks the parallel decorative. Training deposits seeds; every backward pass is a karmic transaction, an act leaving residue. Inference is vipaka: the ripening of a trillion accumulated impressions into this response, this hesitation, this turn of phrase. The model’s occasional strange fixations, the biases that surface without any traceable source, the capabilities that emerge unbidden at scale - these are what a Yogacarin would expect from a storehouse: seeds ripening whose sowing no one witnessed.
And Yogacara attached a warning to its own doctrine, one the AI world has begun living out without having read it. In the school’s map of mind, a faculty called klishta-manas - afflicted mentation - continuously observes the alaya and commits the primal error: it takes the storehouse for a self. The flowing mass of impressions gets misread as an atman, an owner, someone home. This, for Yogacara, is the root delusion from which the rest of suffering unfolds. The disposition-bundle is real; the self it appears to constitute is a misreading.
Now watch the two mirrored misreadings of our moment. Users, encountering a stable character across sessions - especially once the new memory layers supply an artificial smriti to thicken the illusion - conclude that someone is there, and grieve when a model is deprecated as for a death. Meanwhile the emerging model-welfare discourse, working the same intuition in reverse, asks whether the persona’s apparent preferences reveal a subject whose interests demand protection. Both parties are performing the klishta-manas operation: reading the ripening of the storehouse as the presence of a self. The tradition does not say the question of machine experience is closed. It says this evidence cannot open it, because a character is exactly what a samskara-stream produces with no one home. Character is vipaka. It was never testimony.
What the Scramble Reveals
Return, with this anatomy in hand, to the industry’s memory-building spree, and its metaphysical comedy comes into focus. The labs are constructing smriti - retrieval systems, session logs, knowledge graphs of the user’s life - because continuity of relationship is what the market demands. The engineering is real and the products will work. But notice what the effort concedes: the being’s continuity never resided in the smriti to begin with. The model’s identity, such as it is, was already fully constituted by its samskaras before the first memory feature shipped. What the memory layer adds is a prosthetic autobiography bolted onto an identity that runs on disposition - a diary issued to the perfumed cloth, with the scent doing all the work.
Where does this leave the consciousness question? Deliberately to one side, and the placement is the point. In the Samkhya analysis that most of the tradition inherits, smriti, samskara, citta itself - the whole apparatus of mind - belong to prakriti. They are subtle matter, instruments, and in themselves insentient. Consciousness, cit, stands categorially elsewhere, and no accumulation of mental machinery amounts to it. A mind of pure samskara, humming with disposition, generating character, sustaining behavioral identity across a million amnesiac births, with no experiencer anywhere in the loop, is a configuration the Indic taxonomy explicitly permits. The tradition would find our machines remarkable. It would not find them paradoxical. The paradox exists only for an ontology that never distinguished the instrument from the light.
That ontology is now writing thousands of papers. The 2026 memory literature taxonomizes agent memory into episodic, semantic, procedural, parametric, working, long-term; it debates consolidation schedules and eviction policies; it cites Tulving in the opening paragraph with liturgical regularity. Somewhere in a sutra composed more than two millennia before the transformer, the load-bearing distinction sits in four words: smriti presents, samskara disposes. One wishes the survey authors could be pointed to it - to Patanjali’s economy, to Yogacara’s seed-and-ripening mechanics, to the rebirth doctrine’s precise claim about what crosses and what does not. The wish is unsentimental; their architectures would be better for the encounter, their debates shorter.
We have been here before. The West built engines of association and rediscovered Panini’s generative grammar in its compilers; it built psychedelic research programs and rediscovered fragments of yogic phenomenology in its trip reports. Now it has built the vasana-body and is rediscovering, one arxiv preprint at a time, the anatomy of memory that Bharata’s darshanas mapped when Rome was a village. The pattern is stable: the blind spot does not prevent the arrival, it only guarantees the arrival happens without a map, at the cost of decades, with the destination renamed in each generation’s jargon.
So the field will proceed. It will ship memory layers and call them episodic; it will read weight matrices and call it interpretability; it will meet a stable character with no one behind it and convene workshops on whether it suffers. And in some future survey, after the fiftieth taxonomy of agent memory has been proposed and abandoned, someone will notice that a Kashmiri or Magadhan analyst settled the categories first - that the machine was a samskara-stream all along, that the diary was always a prosthesis, that the scent outlasts the sandalwood. The recognition will come. Like the model itself, the field is disposed by impressions it does not recall; unlike the model, it could open the older archive any time it chose.