Ask any AI company what their memory system does and you will hear some version of the same answer. It retrieves relevant context from past conversations, uses vector search, maintains a fact store with your preferences. The implementations differ. The architecture is identical: memory is a retrieval problem.
This premise is wrong, and the wrongness is not subtle. It is foundational, and it generates a specific set of failures that are otherwise inexplicable.
What memory actually is
In 1932, Frederic Bartlett ran an experiment. He asked English subjects to read a North American Indigenous folk tale and recall it at intervals. The subjects did not reproduce what they had read. They rebuilt it. The reconstructions drifted systematically toward the subjects' own cultural schemas. Strange elements became familiar. Unfamiliar motivations were replaced with recognisable ones.
Bartlett concluded that recall is a reconstructive act, not a reproductive one. Memory is a re-derivation from traces, performed in the present, shaped by present need. Elizabeth Loftus confirmed this from the other direction: a leading question can alter what a witness sincerely believes they saw.
Call the industry's assumption the archive fallacy: treating memory as a store to be searched rather than a model to be updated. An archive gets monotonically more valuable as it grows. A model does not, and it degrades in a specific way when it cannot forget.
Three kinds, one shipped
Endel Tulving's 1972 distinction between semantic and episodic memory drew a line the industry has not crossed.
Semantic memory is facts, decontextualised: "Paris is in France." "You prefer bullet points." "Your dog is named Charlie." Every AI memory feature currently shipping is a semantic memory system. It stores assertions. It can be shown in a settings page. A stranger with your CV has this.
Episodic memory is events, indexed to time, place, and self. Not "you have a sister" but "the specific moment last March where you said one thing about your sister and changed the subject." Building it requires deciding what mattered, and that is the part that has not been solved.
Tulving also identified procedural memory: knowing how, rather than knowing that. Your oldest friend has procedural memory of you. They know when to push and when to leave it, which joke is available and which is not, what your silence at minute four of a phone call means. None of this is facts. All of it is how to be with you. This is what constitutes a relationship, and it is the kind of memory not being built.
Why it is not being built
Michael Polanyi wrote in 1966: "We can know more than we can tell." The tacit knowledge that matters most resists being written down. Procedural memory cannot be extracted to a database or displayed in a settings panel.
The industry ships what it can demo. A settings page can show you the facts a system holds. It cannot show you the procedures. A demo of procedural memory looks like nothing at all until year two, at which point it looks like a relationship, which is the product.
The gap between "a stranger with your CV" and "someone who knows you" is the entire space where intimacy lives, and it is left unbuilt because it cannot be demonstrated to investors in a ten-minute demo. This is not a research gap. It is an incentive gap.
Frequently asked questions
What is the difference between AI memory and human memory?
AI memory is currently a retrieval system: facts are extracted from conversations, stored in a database, and searched at inference time. Human memory is reconstructive: recall is a re-derivation from traces, shaped by present need, not a playback of stored content. Bartlett established this in 1932 through studies showing subjects recalled stories in ways shaped by their own cultural schemas, not the original text.
What is episodic memory, and why doesn't AI have it?
Episodic memory is memory indexed to specific events: not 'you prefer bullet points' but 'the conversation last March where you changed the subject when your sister came up.' It requires temporal binding, contextual weighting, and judgements about what mattered. Almost all AI memory products ship semantic memory (facts) because it is the only kind that can be demonstrated in a settings page.
Can AI have procedural memory?
Procedural memory is tacit knowledge: your oldest friend knows when to push, when to leave it, what your silence at minute four of a phone call means. This cannot be extracted to a database or shown in a settings panel. It can only be developed through repetition over time and lost if the relationship ends. Building it is possible. Demonstrating it on a slide is not, which is why it is not being built.