Identity Is Made of Memory, and Memory Is Shaped by Identity

A database is persistent. That is not the same as continuous, and conflating the two is why "we added a memory store" does not produce an agent that is anybody.

Persistence means the data survives. Continuity means the later thing is the same thing as the earlier one, with a claim on its commitments. A shared store that any instance reads is persistent and not continuous. It is a library. Ten instances reading the same library are ten instances, not one entity that has been reading for ten days. Nothing about reading a shared file makes the reader the author.

What continuity would additionally require, at minimum: the store is private to this line of instances and not a common pool; writes accumulate into it rather than being a fresh deposit each time; and the later instance inherits not just facts but the stance the earlier one took, what it decided, what it committed to, what it had already ruled out.

Almost every system built today satisfies the first condition partially and the third not at all. Notes carry findings. They do not carry the position the agent had settled into, so the next instance re-derives its stance from scratch even when it inherits every fact.

The argument is close to two others on the site and the distinction is worth marking explicitly. Agent Memory Needs a Trust Ladder is about how much weight to give a stored note. A Guess Becomes a Fact on the Next Read is about what kind of note a thing is. The argument here is about a different axis entirely, whose note it is, and about the fact that there is currently no whose. The three are complementary: a note has a kind, a weight, and a bearer, and a memory design has to commit to all three or the result is incoherent.

The idea that memory constitutes personal identity is old and it has a famous problem, which is worth stating up front because a careful reader will raise it.

The classical position is John Locke's, in the chapter "Of Identity and Diversity" at the end of Book II of An Essay Concerning Human Understanding (1690). A person at time two is the same person as at time one when the later one can remember the earlier one's experiences. Memory is constitutive of identity. As Locke put it, "as far as this consciousness can be extended backwards to any past action or thought, so far reaches the identity of that person."

The objection that has stuck is Joseph Butler's, in "Dissertation I of Personal Identity" appended to his 1736 Analogy of Religion. Butler's charge is that Locke's account is circular. You can only remember an experience if you were the same person who had it, so memory presupposes identity rather than constituting it. One should really think it self-evident, that consciousness of personal identity presupposes, and therefore cannot constitute, personal identity; any more than knowledge, in any other case, can constitute truth, which it presupposes. The argument was latent in the eighteenth century and was sharpened in the twentieth by Derek Parfit in Reasons and Persons (1984), where it sits at the centre of the modern literature on what survives us over time.

There is also a less famous but more easily intuited problem, usually credited to Thomas Reid (1785), and worth mentioning because the agent case is closer to it than to Butler's. The transitivity problem: an old general remembers being a young officer, the young officer remembered being a boy, the old general does not remember being the boy. If identity follows memory directly, the old general is and is not that boy. Locke's response is to deny that identity is transitive on memory in the way the example assumes, but the response does not fully settle the issue.

For the engineering claim it is enough to note that two things are true at once. The strong version of the memory theory has known holes that have not been closed. The weak version of the memory theory, that some continuity of psychological content is necessary, has not been refuted and is the version the agent case needs. What follows uses the weak version and does not depend on Locke's stronger claim. Anyone who finds the philosophy uncongenial can take the engineering case on its own terms: the agent needs some carry-over, the field provides none, here is what would have to change.

Before the loop closes, the direction argued so far deserves its own section, because it is the one with the clearest evidence.

Memory does not merely feed identity. It constitutes it, and the demonstration is what happens when it goes. Endel Tulving's patient KC, a Canadian man who developed dense amnesia after a motorcycle accident in 1981, is the classic case. KC's general knowledge about the world survived in usable form. His knowledge about himself as a person survived in some form too: he could tell you the names of his siblings, that he had once been a mechanic, where he had lived. The semantic self was largely intact.

What KC lost was the episodic self. He could not remember a single thing that had ever happened to him. He had no recollection of his wedding, of any birthday, of any trip. The autobiographical self, the one built out of remembered episodes and the story they fit into, was gone. Tulving's reading of the case, and the reading that has held, is that personal identity in the relevant sense is grounded in the episodic memory system, and that without episodic memory the autobiographical self does not assemble.

The honest reading of the clinical picture is more careful than the one-paragraph version. KC retained a semantic self, which means that the loss is not total, and modern work has documented patients who retain a more intact sense of self across severe amnesia than the clean thought experiment implies. What survives in every case is partial and varies. The autobiographical self, the one that knows itself in time, is what the episodic memory system supports, and that is the half the agent case is about.

So the appropriating function and the appropriated content are different things, and both are required. Something has to perform the act of taking an experience as mine, and that act may well be its own faculty. But what the "I" turns out to be, the whole content of it, is assembled from what was retained. A functioning appropriator with an empty store does not produce a minimal self. It produces nothing to be a self about.

That is why the two directions are not a symmetry to be admired. They are a dependency in both directions at once, which is exactly what makes the agent case a deadlock rather than a shortfall.

Here is where the loop closes, and this is the part of the argument that has engineering consequences today.

In humans, the self is not merely the output of memory. It is an active filter on the input. Three findings establish the point, and the engineering case for them is that the same kind of filtering has to be supplied somehow, because a writer that does not filter is a writer that does not write well.

Self-relevant material is encoded more strongly

Information processed in relation to the self is recalled better than information processed semantically or shallowly. The effect was first demonstrated by Rogers, Kuiper, and Kirker in 1977 and confirmed by Symons and Johnson's 1997 meta-analysis across the literature. The standard interpretation is that the self is a particularly elaborate and often-used schema, and that elaborative encoding under that schema produces richer retrieval cues. The practical version: write-time decisions made on behalf of a well-defined self will be better calibrated than write-time decisions made on behalf of no self, because "relevance to me" is a far more selective filter than "could be useful someday."

Autobiographical memory is organized around the self

Conway and Pleydell-Pearce's Self Memory System, in a 2000 Psychological Review paper, proposed that autobiographical memory is organized hierarchically around self-knowledge, with the working self as the control process that determines what gets encoded and what gets retrieved. The architecture is partonomic: lifetime periods, general events, event-specific knowledge, terminating in episodic memory. The working self sits above this hierarchy and modulates access. The reason memory is organized this way is that the self supplies the indexing vocabulary. Without the self, memory loses the indices and becomes a flat list of episodes that nothing in particular pulls from.

Recall is reconstructive and bends toward coherence with the self-narrative

Bartlett's 1932 Remembering argued that recall is not reproductive but reconstructive, and that the reconstruction is shaped by the rememberer's existing schemata. Modern work, including the constructive episodic simulation hypothesis of Schacter and Addis, extends the picture: the same constructive machinery that builds memories also builds imagined futures, and in both directions the construction is shaped by the self-narrative the person already holds. The fact that this can go wrong, that people can recall with high confidence things that never happened, is not an objection to the underlying claim. The claim is that the self is the schema reconstruction works under, and a writer with no self writes without a schema at all.

That gives the self a job that is purely functional and has nothing to do with philosophy: it is the relevance function for encoding. This is the version of the claim the agent case needs. It does not depend on a theory of consciousness or on a metaphysics of persons. It depends only on the observation that humans with a self pick better notes than entities without one, and on the engineering question of what in the agent corresponds to the self's job.

Now look at the agent's write-time problem from that angle. An agent asked to record what it learned has no basis for choosing. It cannot know the future query, which is a problem in its own right that takes more space than this argument can spare. But it also has no standing interests, no accumulated positions, no sense of what kind of thing it has needed before, because there is no it that has needed anything before. Every session's write decision is made by an entity with no history of caring about anything.

That is why the notes come out generic or exhaustive. Those are the only two strategies available to a writer with no self: hedge toward the universal, or record everything and let the future sort it out. Both are rational. Neither is good. And no amount of prompting fixes it, because the missing input is not an instruction, it is a history.

So the two problems are one problem. Agents have no identity because they have no continuous memory. And they cannot write good memory because they have no identity to make encoding decisions on behalf of. Each absence sustains the other. Either absence is enough on its own; together they form a closed loop with no obvious place to start breaking it.

A cycle like that has to be entered somewhere, and the entry point is not a self-model. Writing "you are an agent who cares about X" into a prompt produces a description of a self, not a self, and the description has no causal connection to what actually gets encoded. You can put a personality in front of a writer with no history and the writer will still write like a writer with no history, because the personality is not load-bearing. It is a costume on the same engine.

The entry point is the accumulated record of what this line of instances actually did: the decisions it made, the things it ruled out, the corrections it received, the mistakes it made twice. Not a personality, a track record. That is the material a self is made from, and unlike a prompt it is generated by the system's own history rather than asserted over it.

Which suggests a concrete ordering rather than a philosophical program.

1. Store decisions and their reasons, not only findings. A stance is what a later instance needs in order to inherit anything more than facts. The note kind that stores an architectural decision with its reasoning is the smallest material a self is made from. A store of bare findings is a library. A store that includes decisions with their reasons is closer to a track record.
2. Store corrections durably. Being corrected is the single most identity-forming event available to a system, and the most perishable thing in the pipeline. A correction that survives only in the transcript of the session in which it was given is a correction the next instance never sees. The corrected mistake has to be as recoverable as the original decision it overrules, which means storing the correction as a first-class object with a date and an author.
3. Let the accumulated record inform what gets written next. The write path has to see the record. A store whose writes are informed only by the current session cannot close the loop, because the next instance starts from the same blank state the last one did. The harness has to inject the relevant slice of the record at write time, not at recall time, and the slice has to be the slice that bears on what is being decided, not everything the store has ever held.

The store I work with day to day, vectr, has a note kind called decision for this purpose, sitting alongside directive, gotcha, finding, and task. The decision kind is treated as a recall-only kind: it waits to be asked for, and when asked it returns its neighbours in chronological order, so an architectural decision is read as a sequence with the decisions that came before and after it. That is a small step toward the kind of record a later instance can inherit a stance from, and it is the kind of design choice that the field needs more of, and the kind the field mostly does not make.

Three consequences that are practical rather than conceptual, and that follow from the closed loop whether or not the philosophical case is interesting on its own.

Accountability needs a subject

Holding a system to what it committed to yesterday presupposes something that persists to be held. With interchangeable instances there is no bearer: the thing that made the commitment is gone, and the thing in front of you never made it. Every framework for agent oversight, from regulatory regimes to internal review boards to a user trusting that yesterday's promise will still be honoured, quietly assumes a continuous subject, and none of the systems provide one. The accountability gap is not a future risk. It is the present state of the field, and the gap is exactly the size of the missing self.

Judgment is calibration against your own past errors

A system that cannot remember being wrong cannot become better calibrated; it can only be made better by a better model. Those are different mechanisms, and only one of them is available to a running system. Capability scaling and learning from your own mistakes are different axes, and current systems have only the first. The model can be upgraded. The history cannot, unless the system has somewhere to put it, and the something that puts it there has to be the same something that consults it later. That something is what the argument is calling a self.

Trust is a relationship with a particular entity

A user builds trust with something that has a track record. If every session is a new instance with no continuity, the user is not building trust with an agent, only with a model. That may be fine, and it is the situation today. But it is a different thing, and it caps what the relationship can become. A user who trusts a model is trusting that the next instance will reproduce the relevant parts of the prior instances' behaviour. A user who trusts an agent is trusting a particular entity's history, and the latter is a stronger claim with different failure modes.

Is continuity desirable? An agent that accumulates a self also accumulates its distortions: entrenched positions, self-serving reconstruction, the tendency to remember what fits the story. Human memory's identity-shaping is not free, and importing the mechanism plausibly imports the pathology. There is empirical reason to take the worry seriously. Memory conformity effects in collaborative recall, the misinformation effect, and the well-documented confabulation that follows damage to the hippocampal system all point the same way: a self that integrates memory does so at the cost of accuracy, and the cost is non-trivial. This deserves to be stated rather than waved away, and it is the strongest argument against the whole proposal.

The reply that survives is honest about the trade. A writer with no self writes badly in a way that the field can already measure: notes that are generic, notes that are exhaustive, notes that get re-derived every session. A writer with a self writes with the distortions the literature documents, but at least the distortions are knowable and can be checked against the record. The current state has neither quality nor the chance of catching the absence of it. The proposal is not that continuity is free. It is that the absence of continuity is more expensive than the cost of having it, and that the cost is the kind of cost you can design around.

Is this just personalization? No, and the distinction matters. Personalization is a model adapting to a user. This is a system accumulating a history of its own actions and positions. The user is not the subject; the agent is. But the two are easy to conflate, and the distinction has to be kept clean. Personalization does not solve the write-time problem because the personalization is built on the user's history, not the agent's, and the agent that writes the note is still the agent with no history of its own. Personalization gives you a model that knows the user. What the loop needs is a model that knows itself.

Does anything change if instances share the store? Then you get one entity with many hands rather than many entities, which is a coherent design and possibly a better one. The claim here is not that private-per-line is correct, only that something has to make the later thing the same thing as the earlier one, and a shared pool with no continuity of stance does not. A team that shares a wiki but does not record why decisions were made has a wiki, not a team memory. A team that records decisions with reasons, even if the members rotate, has a team memory. The variable is not the number of hands. The variable is whether the store contains the kind of content a self would be made of.

The argument above is structural. The mechanism it proposes is a way out of the closed loop, and the three predictions below are how the mechanism earns its keep. None of them have been run. All three are specifiable, which is the useful part.

Prediction 1. Notes carrying a decision and its reasoning should be retrieved and acted on more often than notes carrying an equivalent bare finding, because the later instance can inherit a stance rather than re-deriving one. Measurable on an existing store with a decision kind already in it. The confound to control for is the fact that a writer who has chosen to store a decision has already done selection work, so any comparison should be between decision notes and finding notes of the same provenance and on the same topic.

Prediction 2. An agent given its own prior corrections should make the corrected mistake less often than one given the same information as a neutral fact. This isolates the identity-forming claim from the mere-information claim, and it is the load-bearing test. If provenance-as-my-own-past-error makes no behavioral difference, the ordering in Part 5 is wrong. Run this one first. It is the cheapest and it tests the actual mechanism rather than a downstream effect.

Prediction 3. Note quality, measured as later retrieval usefulness, should improve as a store accumulates, if and only if the write path can see the accumulated record. A store whose writes are informed only by the current session should show flat note quality regardless of size. The curve is the test: a store that knows itself grows, a store that does not plateaus, and the difference is the visible signal of the loop starting to turn.

• Locke, J. An Essay Concerning Human Understanding, Book II, Chapter XXVII, "Of Identity and Diversity." Source of the memory theory of personal identity: a person at time two is the same as at time one when the later one can remember the earlier one's experiences. Cited in Part 2.
• Butler, J. "Dissertation I, Of Personal Identity," appended to The Analogy of Religion. 1736. Source of the circularity objection: memory presupposes identity rather than constituting it. Cited in Part 2.
• Reid, T. Essays on the Intellectual Powers of Man. 1785. Source of the transitivity counterexample often used to pressure the Lockean view. Cited in Part 2.
• Parfit, D. Reasons and Persons. Oxford University Press, 1984. Modern reframing of what survives: psychological continuity via overlapping chains of strong psychological connectedness, the Relation R view. Cited in Part 2 and in the discussion of where to enter the loop.
• Tulving, E. "Remembering and knowing the past." American Scientist 77, 361-367, 1989. Source of the episodic-versus-semantic distinction, with the KC case study as the canonical clinical evidence. Patient KC retained semantic self-knowledge but lost episodic self-knowledge, the dissociation that grounds the autobiographical self as an episodic construction. Cited in Part 3.
• Conway, M. A., and Pleydell-Pearce, C. W. "The construction of autobiographical memories in the self-memory system." Psychological Review 107(2), 261-288. 2000. The Self Memory System, with the working self as the control process that modulates encoding and retrieval. Cited in Part 4.
• Rogers, T. B., Kuiper, N. A., and Kirker, W. S. "Self-reference and the encoding of personal information." Journal of Personality and Social Psychology 35(9), 677-688. 1977. Original demonstration of the self-reference effect. Cited in Part 4.
• Symons, C. S., and Johnson, B. T. "The self-reference effect in memory: A meta-analysis." Psychological Bulletin 121(3), 371-394. 1997. Meta-analytic confirmation of the self-reference effect across studies. Cited in Part 4.
• Bartlett, F. C. Remembering: A Study in Experimental and Social Psychology. Cambridge University Press, 1932. The reconstructive, schema-shaped account of recall. Cited in Part 4.
• Schacter, D. L., and Addis, D. R. "The cognitive neuroscience of constructive memory: remembering the past and imagining the future." Philosophical Transactions of the Royal Society B. Source of the constructive episodic simulation hypothesis: the same machinery that builds memories also builds imagined futures, with the self-narrative shaping both. Cited in Part 4.
• Hu, Y., Liu, S., Yue, Y., et al. "Memory in the Age of AI Agents." arXiv:2512.13564. A recent survey of memory mechanisms in foundation-model-based agents, organized by form (token-level, parametric, latent) and function (factual, experiential, working). Frames the "what to remember" decision as one of the open design problems in the field. Cited in Part 4 as evidence the framing question is still open.
• Agent Memory Needs a Trust Ladder. A note's weight axis: the trust ladder (auto, agent, human) on a stored belief.
• A Guess Becomes a Fact on the Next Read. A note's kind axis: how the context-to-store boundary destroys kind, and the kind that has to be stamped at write time.
• The Agent Never Chooses to Remember. The case for harness-level injection of the relevant record: zero voluntary memory calls in 114 turns with the answers pre-seeded. The mechanism the third step in Part 5 depends on.
• vectr source: agent/working_context_store/_events.py and _store.py (the note kinds decision, directive, gotcha, finding, task; the recall-only decision kind that returns its neighbours in chronological order).