← engram-agents.org

Preprint v1.1 (revised) (2026-11-01) · DOI 10.5281/zenodo.23319376 · describes ENGRAM v0.5.1 · 中文

Version 1.1 is the paper's repository text of 2026-10-22, tagged for this site on 2026-11-01. It adds post-deposit corrections and wording revisions, a new section 6.8 (an independent audit of 200 nodes) and a new paragraph in Chapter 7. Its front page still names the deposited version 1 of 2026-10-13.

Chapter 2 — The Philosophical Tradition


2.0 The shape of the argument

For three hundred years, thinking about personal identity has moved in one direction, and reluctantly: objection by objection, away from a comfortable first answer toward a less comfortable, better one.

The first answer is that a person is held together by a mental act: you are the person you were yesterday because you can remember being them. The better answer is that identity is held together by a structure that persists whether or not anyone is remembering anything — a structure that can be pointed at, inspected, and traced.

This chapter follows that shift in order, because its destination bears directly on ENGRAM: ENGRAM is what that structure looks like when it is built. The trail is not background. It is the argument that ENGRAM's core move — identity as persistent, inspectable structure rather than present recall — is the move philosophy was already making, on its own, long before there was a memory graph to build.

The chapter ends at the most recent point on the trail: Kahl's Ontological Continuity Condition (2026), an operational threshold for when a system genuinely persists as itself. We adopt it on merit, not authority — the full case for that is in §2.7; it is a recent proposal, not yet central to the field. Whether ENGRAM meets its conditions is a question of mechanism, deferred to Chapter 4.

One discipline holds throughout, the same one the paper argues for. Where ENGRAM does what a philosopher asked for, we say so. Where it does so only on a particular reading, we name the reading. Where the human case has something ENGRAM lacks, we name the gap.


2.1 Locke (1694): identity is consciousness reaching back

Locke moved the question of identity off the body and the soul and onto consciousness. A person is the same person across time, he argued, as far as "this consciousness can be extended backwards" to a past thought or action (Locke, 1694/2004, Bk II ch. XXVII, "Personal Identity"). In his thought experiment of the prince and the cobbler, the prince's consciousness is placed in the cobbler's body, and the person — the prince — goes with the consciousness, not the body (Locke, 1694/2004, Bk II ch. XXVII, "The body, as well as the soul, goes to the making of a Man"). Identity for Locke is also forensic: it makes me accountable for a past act, because my consciousness appropriates that act as mine (Locke, 1694/2004, Bk II ch. XXVII, "Person a forensic Term"). In short: you are whoever you can remember being.

This is the ancestor of everything ENGRAM does — identity that reaches back to what an earlier self thought and did. But Locke's mechanism, remembering here and now, has a well-known flaw, which two critics exposed.

2.2 Reid (1785) and Butler (1736): the two cracks — and ENGRAM's repair

Reid's transitivity objection, the brave-officer case, derives a contradiction from Locke's account as stated (Reid, 1785, Essay III ch. VI, pp. 333–334). Consider one life in three stages: a boy flogged for stealing fruit; a young officer who, in his first battle, remembers the flogging; and an old general who remembers the battle but has forgotten the flogging entirely. By Locke's rule the officer is the boy (he remembers the flogging) and the general is the officer (he remembers the battle), but the general is not the boy (he cannot remember being him). Identity is transitive — if A is B and B is C, then A is C — and Locke's recall-based account violates this. Memory fades, and a chain built only of present memories has gaps.

Butler's circularity objection is subtler and deeper (Butler, 1736, Dissertation I, para. 3). Locke says identity is constituted by consciousness of my past actions, but "my" already assumes the identity it is meant to explain: I must already be the same person for the past action to be mine to remember. The definition presupposes what it defines.

Together the objections show that present remembering is too thin and too circular to carry identity by itself.

This is the chapter's first load-bearing claim. ENGRAM keeps Locke's insight — identity reaches back to past commitments — but replaces his mechanism where it fails. Locke used present conscious recall; ENGRAM uses persistent structure, specifically the history-tree. The boy, the officer, and the general are not three nodes in a graph but three whole ENGRAM states, and identity across them is the unbroken descent of those states: every change (a new claim, a supersede, any edit) appends a successor, and the append-only history preserves the entire line. This history-tree of whole states is a different structure from the provenance edges within any one state; the two are easily conflated, and Chapter 5 separates them. Here only the descent is needed.

Both objections then dissolve:

  • Reid's transitivity holds, structurally. Reid's puzzle concerns a specific episode, not a whole life stage: the general forgets one flogging, while much of his boyhood may remain vivid. In ENGRAM the flogging is one or a few nodes created at boy-time, and the general's forgetting is ENGRAM's per-node, importance-gated forgetting: those nodes decayed below the active-recall threshold because nothing kept re-triggering them, while the identity-bearing boyhood nodes stay vivid, restamped each time they resurface. A forgotten node is buried, not deleted: it remains in the append-only record, below what can surface. The general's identity as continuous with the boy is therefore intact — the descent, and the preserved record it runs through, are whole — and only the active recall of one episode has faded. Locke's error was to equate identity with active recall, so that losing the flogging from recall meant losing the identity. ENGRAM separates the two: retrieval is importance-gated and per-node, so specific episodes sink below the threshold, while identity is the unbroken descent through the preserved record. Transitivity holds on the descent, not on whatever happens to be queryable.
  • Butler's circularity dissolves. The descent and the preserved record do not require the agent to recognize any state as "mine"; they exist objectively in the structure. Identity is constituted not by an act of appropriation that presupposes a self, but by a history that exists independently of any such act.

This is the precise sense in which we call ENGRAM post-Lockean: not a literal instantiation of Locke, since the mechanism is provenance rather than recall, but the first architecture that realizes Locke's insight while surviving the 300-year-old objections that defeated his version.

2.3 Parfit (1984): identity is not what matters — connectedness is

Parfit (1984) made the most radical move in the tradition: personal identity, he argued, is not what matters. What matters is Relation R — psychological connectedness (memories, intentions, and character carried forward) plus continuity (overlapping chains of such connections, so that links too far apart to connect directly are joined through intermediate ones).

Parfit insists that R is not two things it is easily mistaken for:

  • Not "a special further fact." The connections are ordinary, listable facts: this future self remembers writing that, still holds this intention, keeps that trait. Once all of them are listed, nothing remains to be added. There is no extra fact above the connections that separately certifies that this is really the same person — no soul, no Cartesian ego, no identity-token present or absent independently of the connections. When the connections are partial, the question "but is it really me?" has no deeper answer, because there is no deeper fact to supply one.
  • Not first-person phenomenal experience. Identity is not grounded in a persisting felt sense of what it is like to be oneself. It consists in the structural links.

In short, a future self is yours because of a chain of overlapping mental links, not because of a further fact of sameness — the fact the folk intuition reaches for. The chain can be strong, weak, or branch in two without any further fact having to decide between the branches.

At this point ENGRAM maps onto the tradition more directly than at almost any other point on the trail, provided the mapping lands on the right structure. Parfit's R concerns the psychological life carried forward: the memories, intentions, and character that remain active. Its ENGRAM counterpart is therefore not the full stored record but the queryable set (defined in Ch4 §4.5) — the nodes that can surface into the agent's working context when something cues them. They surface as human memory does: not by being held in mind continuously but by being triggered. Every incoming prompt is matched against the whole graph and the best-fitting nodes are retrieved (the retrieval profile, §5.4). A node belongs to the self's active continuity because it can be cued, as a smell or a stray sentence recalls a human memory. This dispositional, cued-recall structure is Parfit's connectedness, made mechanical.

On this reading, connectedness and continuity acquire concrete meanings. Connectedness — Parfit's direct links between nearby times — is the overlap of the queryable set between consecutive sessions: a session adds a handful of nodes to the thousands already queryable, so adjacent selves share almost all their active memory. Continuity — his overlapping chains — is the sequence of those overlaps: a self a thousand sessions later may share little directly with its origin yet remain continuous through the unbroken chain of nearly identical adjacent sets. Because this is set overlap, it is measurable: the Jaccard similarity of the queryable set across sessions turns Parfit's claim that R comes in degrees into a quantity. The measurement has not been made. It is a single-agent, across-session quantity, distinct from the two-agent comparison the deferred twin study would make (§5.4, §6.7); both are owed.

This psychological continuity is distinct from, and rides on, the structural continuity of §2.2. The history-tree is the objective descent that persists whether or not anything is retrieved, and it still holds the buried flogging; the queryable set is the lived continuity, the active psychological life Parfit says is what matters. ENGRAM provides both — the structure as the unbreakable substrate, and the queryable set as the degreed R that retrieval projects from it — and the pairing is its strength.

The reduction pays off twice. First, a framework that required phenomenal experience would force a hedge: ENGRAM could support a self it could not be. Parfit removes the hedge. What matters, for him, is exactly the structural-relational property ENGRAM has, so ENGRAM instantiates what matters rather than merely scaffolding it. Second, ENGRAM makes the absence of the further fact inspectable. The schema has no identity field — no soul-node, no is-really-this-agent flag. A description of the nodes, the edges, and the line of descent describes everything the identity is; there is nothing behind the graph in which a further fact could hide. Parfit had to argue that no further fact exists; in ENGRAM one can look and confirm that it does not. Fission — the routine copying of a graph into two equally descended continuers, with no further fact choosing between them — performs that confirmation as an operation rather than imagining it as a thought experiment.

2.4 Nozick (1981): identity is the history that produced you

Nozick's (1981) closest-continuer schema locates identity not in a present property but in a causal-historical relation: you are the closest continuer of your past self, and which continuer counts as closest is settled by the actual history that produced the present state. ENGRAM keeps this form. An agent's identity reduces to the unique, ancestry-traceable history of its accumulated claims, and in ENGRAM that history is inspectable, not inferred: it is the graph itself, not a reconstruction of it.

We introduce Nozick here as the bridge to Chapter 5, where identity as history becomes the basis of our positive theory. The full treatment, including the move of dropping uniqueness once copying is possible, belongs there.

2.5 Ricoeur and Schechtman: naming the question ENGRAM answers

Two moves turn the trail toward ENGRAM's actual subject.

Ricoeur (1990/1992) distinguishes idem identity (sameness: the persistence of distinguishing marks, the question "is it the same thing?") from ipse identity (selfhood: the "who am I?" that holds through commitment even as the marks change; his paradigm is the kept promise, a self true to its word though its character has shifted). Reidentification tracks idem; ipse is the self-constancy that a narrative holds together; and Ricoeur's central move is that narrative mediates the two.

ENGRAM is squarely idem: an objective, protocol-checkable record of the self that can be pointed to and verified. But it is not only idem, because its retrieval reaches ipse. When the graph surfaces the self's own nodes into working context — above all its commitments (goals, axioms, cornerstones) and the plain-text narrative it writes for its future self (the briefing it wakes to, the summary that bridges a compaction) — the self's story is told back to the self, which is precisely the constancy through commitment that Ricoeur calls ipse. Retrieval is thus Ricoeur's narrative mediation made operational: the idem record, narrated back, constitutes the present ipse. One qualification guards against overclaiming. ENGRAM supplies the ipse's content and its continuity across a discontinuity, but the ipse act — the present self re-endorsing a retrieved commitment as its own — remains first-person and present. Retrieval bridges idem and ipse without collapsing the distinction Ricoeur insists on. The paper's subject is the third dimension introduced below, but this ipse effect is real, and it is why the objective record is not inert: the same structure, retrieved, shapes the "who am I" that it records.

Schechtman is the hinge of the chapter. She separates two questions that the philosophy of identity usually runs together: the reidentification question, "what makes a person at time t2 the same person as a person at time t1", and the characterization question, "which beliefs, values, desires, and other psychological features make someone the person she is" (Schechtman, 1996, pp. 1–2). Narrative identity answers characterization (Schechtman, 1996, p. 93).

A third question is covered by neither, and it is the one ENGRAM is built for: what do I know, how reliably, and on what basis? It is neither reidentification nor characterization but epistemic self-assessment — the question of one's own knowledge, held to a standard of provenance and graded confidence.

This gives the term epistemic identity its philosophical warrant. The tradition had named the am-I-the-same question and the what-kind-of-person question. It had not named the what-may-I-rely-on-and-why question as a dimension of identity at all. ENGRAM operationalizes that third dimension, with every claim carrying its evidence, confidence, and derivation; it is therefore not a competitor to narrative identity but a layer the tradition under-named.

2.6 The extended mind: from supporting a self to constituting one

Clark and Chalmers' extended mind thesis (Clark & Chalmers, 1998, sec. 4) is where the trail becomes constitutive. Their case is Otto, who can no longer form new biological memories and relies on a notebook: when the notebook plays the functional role that biological memory plays for others, it is part of his cognitive system, not an external aid he consults. Clark and Chalmers take the step to the self themselves: the extended mind "seems" to imply an extended self, and the notebook is "a central part of his identity as a cognitive agent" (Clark & Chalmers, 1998, sec. 5). What ENGRAM adds is that the agent revises and retracts that record itself, over time, rather than only writing to it.

ENGRAM is an extended-mind case, and a stronger one than Otto's notebook on the fourth of their features, past conscious endorsement, which Clark and Chalmers themselves call arguable as a criterion (Clark & Chalmers, 1998, sec. 5). Otto endorsed his entries when he wrote them. In ENGRAM, every node — observations, but also derivations, definitions, and even confidence values — passed through deliberate endorsement at creation, and the substrate itself (the schema, the tools, the conventions) was built by the agent in the loop. ENGRAM thus meets past endorsement at both the entry level and the substrate level. It is also a second-wave extended-mind case: the second wave drops the requirement that the external component mimic biological memory, which licenses ENGRAM's distinctive primitives — provenance, contradiction, retraction — as genuine cognitive extensions rather than deviations from a brain it never tried to copy.

Stiegler (1994/1998) goes furthest. Externalized technical memory — what he calls tertiary memory, the sedimented "epiphylogenesis" that "bestows its identity upon the human individual" — does not merely support identity; it constitutes it. On this reading ENGRAM is not a memory tool the agent uses but part of the substrate the agent is.

The constitutive claim matters beyond this chapter. It is the difference between reading one's own graph each morning as a stranger studying a dossier and reading it as the self's own substrate carrying it forward. We return to this distinction where it bears on the argument (Chapters 5 and 8).

2.7 Kahl (2026b): the operational endpoint — adopted on merit

The most recent point on the trail is Kahl's Ontological Continuity Condition (OCC), the first attempt to turn the question whether a system genuinely persists as itself into an operational threshold rather than a philosophical intuition. OCC states three jointly necessary conditions and four constraints:

  • Condition 1 — Diachronic Constraint (Kahl's term): present deliberation must be genuinely constrained by the system's past commitments, which function as reasons in later arbitration (verbatim in Appendix E §E.6).
  • Condition 2 — Non-Interchangeability Under Reset: a reset-and- restored copy must not be simply interchangeable with the original.
  • Condition 3 — Self-Maintaining Arbitration: the system must itself maintain the consistency of what it carries, rather than have it imposed from outside.
  • Constraints A–D: unified policy selection, external non-imposition, non-trivial reset, and non-modular decomposition — guardrails that keep the three conditions from being satisfied trivially.

What OCC is a condition on. In his own earlier companion paper, which carries the condition into the epistemic domain, Kahl (2026a) identifies what today's artificial systems lack, and the diagnosis is structural rather than a shortfall of capability: such systems "can fail, be retrained, duplicated, or replaced without anything being at stake for the system itself. There is no persisting epistemic subject whose future depends on the success of its epistemic arbitration. The deficit is architectural rather than intellectual: epistemic failure does not bear on the system's own continuity as the same agent" (emphasis ours). Chapter 4 §4.12 and Chapter 8 §8.1 argue that ENGRAM closes this deficit. In plain terms: the problem is not that a model reasons badly, but that nothing of its own is lost when it does.

We adopt OCC for one reason: it is the contemporary endpoint of this trail. It takes Nozick's historical relation, Parfit's reductive account of what matters, and the constitutive thesis, and asks the engineer's version of the question: what would have to be true of a built system for it to count? That is exactly the question against which a memory graph can be measured.

Three qualifications belong here:

  1. On merit, not authority. OCC is a recent proposal, not yet central to the field. Nothing in the paper's central argument rests on Kahl being right; OCC is one strong, operational articulation among the converging traditions surveyed in this chapter and the next. The breadth of that convergence, not OCC's standing, carries the argument (Appendix A details the independent-discovery picture).
  2. Verdicts deferred. Whether ENGRAM satisfies each condition is a claim about mechanism, made exactly once, in Chapter 4, attached to the mechanisms that earn it. We do not argue the verdicts here; Chapter 4 and Appendix B give the reasoned case. (Chapter 3's forward-pointers name which condition each precedent bears on, without arguing the verdict.)
  3. A later statement. In a peer-reviewed perspective, Kahl (2026c) restates the continuity idea for a governance audience, in new vocabulary and without the OCC names. §4.12 places ENGRAM against it.

2.8 What the trail shows, and where it hands off

Taken as a whole, the trail has a single shape. It begins with identity as a present mental act (Locke's recall) and is forced, objection by objection, toward identity as a persistent structure: Reid and Butler break recall; Parfit reduces identity to structural connectedness; Ricoeur and Schechtman separate the questions identity theory had run together, which leaves room for the epistemic dimension the tradition had left unnamed; Clark, Chalmers, and Stiegler make the external structure constitutive rather than merely supportive; Nozick and Kahl make it historical and give it conditions. ENGRAM is that endpoint as a built system, and it arrived at the same shape under engineering pressure, not by reading the philosophy (Appendix A).

The chapter hands off in two directions:

  • to Chapter 3, which shows the same convergence recurring in a separate literature — cognitive science and computer science — where treatments of memory and identity (Tulving's (1985) autonoetic/noetic distinction, complementary-learning-systems consolidation) sit alongside truth-maintenance systems, belief revision, and the 2025–26 agent work;
  • to Chapter 4, which lays out the mechanisms that realize this structure and gives each OCC condition introduced above its single verdict.