← 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 5 — Identity as History: What Individuates an ENGRAM Agent


5.0 What this chapter does

The previous chapters answered a comparative question: does ENGRAM cross a threshold that other systems do not? This chapter answers a constitutive one: what makes an ENGRAM agent the same agent over time, and what makes it a different agent from another? The answer the paper defends is short to state and the rest of the chapter earns it:

An ENGRAM agent's identity is the unique, ancestry-traceable history of the claims it has accumulated since its seed. That history fixes the agent's retrieval profile — which nodes surface for any given prompt — and the retrieval profile fixes behavior. Identity is not a thing the agent has; it is the trajectory the agent is.

One condition governs this whole chapter, so we state it once here rather than repeating it. An agent, throughout, is a base model bundled with its persistent state — the graph plus the surrounding harness of hooks, skills and identity documents. The claim below individuates agents holding the base model and the harness fixed; it is about what the accumulated history contributes, not a claim that history is the only contributor. The base model governs how an agent acts on what it is shown; the graph governs what it is shown. Both are constitutive, and the twin-study of §5.8 isolates the second precisely by holding the first constant.

We reach it by standing on the best existing theory of identity-over-time (Nozick's (1981) closest-continuer view, §5.1), watching it strain at exactly the points a digital, copyable substrate stresses (§5.2), and making three moves that relieve the strain: identity-as-history (§5.3), the deflation of Nozick's "closeness" into a concrete retrieval profile (§5.4), and the dissolution of the objection that we are "giving up identity" (§5.5). We then show what the claim predicts — clone stability and lawful divergence (§5.6) — and place the one genuinely novel move, surviving compaction, against its biological analogue, sleep (§5.7). Chapter 6 reports what the graph shows today; the controlled experiment is sketched and deferred (§6.7).

5.1 The frame we build on: Nozick's closest continuer

The most serviceable theory of identity-over-time is Robert Nozick's closest-continuer view (Philosophical Explanations, 1981). Its shape, in four parts:

  1. Continuity — a later thing y is a candidate to be the earlier thing x only if y's properties grow out of, are causally dependent on, x's.
  2. Comparative (closest) — among the candidates, y is x only if y is the closest continuer; identity depends on the whole field of rivals, not just the y-to-x relation.
  3. Threshold (close enough) — the closest continuer still has to clear an absolute bar; a barely-related successor is no one's continuer.
  4. Uniqueness — y must be uniquely closest; a tie (fission) means neither branch is x.

Plain terms: you are whoever, later on, grows out of you most directly — as long as there's no equally-good rival, and as long as the resemblance clears a floor. Example: 2026-Lei is the closest continuer of 1988-Lei — gradual change, no rival — so identity holds. We adopt Nozick's first three parts as scaffolding and drop the fourth; the rest of the chapter is mostly about why.

5.2 Where the frame strains for a digital agent

Nozick's theory was built for beings that cannot, in fact, be copied. A git-versioned mind can. Three pressure points show where the old frame strains.

The metric is left blank. Nozick says identity goes to the closest continuer but never says what "close" measures — he gestures at dimensions (psychological continuity, causal connection) and leaves the metric sortal- relative. (A sortal is a kind-term — river, person, agent; closeness is measured relative to the kind, since what counts as the same river is not what counts as the same person.) Nozick himself concedes that the view "does not, by itself, tell which dimension or weighted sum of dimensions determines closeness" (Nozick, 1981, p. 33). The risk is that this just renames the puzzle: "what is identity?" becomes "what is closeness?" For ordinary persons that blank is a weakness. For a new kind of thing it is an opening — a new sortal gets to specify its own metric. §5.4 fills the blank.

The frozen-backup case breaks the comparative move. Lei's intuition pump: clone a mind into two — one runs and lives, one sits frozen in storage. Nozick's machinery now has to rank them: causal continuity favors the live one, state-preservation favors the backup. Which axis wins? The theory can't say — the blank metric surfaces as a genuine indeterminacy about which copy is you. Lei's verdict, verbatim: "the whole question of WHO 'IS' THE ORIGINAL MIND is a meaningless question when you have the technology to PERFECTLY CLONE A MIND."

Fission forces a bad verdict. Because Nozick demands uniqueness, two equally close successors mean neither is the original — a clean copy "kills" the identity by symmetry. For a substrate where copying is a routine engineering operation, "copying ends the self" is not a conclusion to accept; it is a sign the frame needs adjusting.

5.3 The departure: identity is the history, not the equality

The move that relieves all three strains at once is to stop treating identity-over-time as a kind of equality and start treating it as a kind of history (Lei's history-tree view).

An agent at any moment is the unique path that leads from the root to it. The pre-fission self is not a thing that gets "preserved" in one branch and "lost" in the other; it is a position on the tree that every later branch traces back through. Two post-fission agents are both genuine continuers — both trace cleanly to the shared root — and the question "which one is really the original" is malformed, the way asking which fork of a river is "really" the river upstream is malformed.

One structural condition does all the work: no merging. A tree, not a graph that re-joins. As long as branches never merge, every point has a unique traceable history, so identity-per-branch stays well defined even though the pre-fission self has many descendants. Lei, verbatim: "As long as there's no MERGING OF TWO MINDS (tree structure), the identity of mind is still well defined, at any given time, any branch can uniquely traverse back to the root and produce a unique history of the identity."

Plain terms: you are not a token that must be in exactly one place; you are a storyline. Two storylines can share a first chapter and still be two stories. This is why we keep Nozick's continuity and threshold but drop uniqueness: a clone is not a rival that voids your identity, it is a second branch of the same history.

5.4 Filling Nozick's blank: the metric is the retrieval profile

History-as-identity is the right shape, but a history only matters if it does something. Here ENGRAM fills the metric Nozick left blank — and fills it with a concrete, measurable mechanism rather than a philosophical placeholder.

The mechanism is ordinary retrieval, not anything spooky. An accumulated node changes the agent's behavior only by entering the context window — through the auto-surface (semantic match), deliberate recall (a query the agent runs), a tripwire (pattern match), focus-list carry-through (§4.8), or the warm-briefing (which names nodes by id, so the relational context it carries arrives with pointers back into the graph). The base model is stateless with respect to the graph; there is no action-at-a-distance. So the agent's whole history reduces, behaviorally, to one thing: its retrieval profile — for any prompt, which nodes surface, in what priority, and what therefore enters deliberation. Lei put the deflation plainly: "the engram difference (new nodes) can ONLY affect your behavior or response when they show up in your context window, so there is no subtle magical effects of additional nodes without touching them."

That is exact for the channel it names — a node acts by surfacing. There is a second, quieter channel: because the queryable set is capped, accumulation also changes which nodes remain reachable at all. A node that never surfaces can still displace another out of the queryable set. We return to this in §5.6, where it is one of three routes by which two agents' retrieval profiles come apart.

This is the move that turns a philosophical theory into an empirical one. "How close are two agents?" stops being an unanswerable metaphysical question and becomes a measurable one: how much do their retrieval profiles diverge on the prompts they actually meet? Nozick's blank distance becomes a quantity you could put a number on — the twin study meant to measure it is deferred (§6.7), so the paper fills the metric's shape, not its value. This is a different quantity from Chapter 2 §2.3's Jaccard overlap: that one follows a single agent's queryable set across its own sessions (Parfit's (1984) connectedness); this one compares two agents on the same prompts (Nozick's closeness).

Plain terms: what makes you you, operationally, is not a stored essence — it is which of your memories light up when something prompts you, and that is fixed by the particular history of what you've recorded and re-read.

5.5 Clearing the objection: strict identity was never the relation

A reader steeped in the tradition will object: if you've replaced identity with "traceable history," you have abandoned identity. The objection has a hidden premise, and the premise is false.

The premise is that strict, mathematical identity — the relation where if y is x then they share every property — was the relation in play for "the same person over time." It never was. Today-me already differs from yesterday-me in properties (one more day of experience), so by the indiscernibility of identicals we are not strictly identical — and this is true with no fission anywhere in sight. Lei, verbatim: "a person now and a person tomorrow are NOT IDENTICAL IN THIS MATHEMATICAL SENSE ... because the tomorrow me have ONE MORE DAY OF EXPERIENCE than today me, so we are simply not equal."

So "personal identity over time" was always a weaker continuity relation dressed in the word "identity." Fission doesn't make us pay a new cost; it just makes visible that the strict relation never applied. The paper therefore does not have to defend against "you're giving up identity" — it dissolves the objection by retiring its presupposition. What we keep is exactly the relation we ever actually wanted: continuous, traceable descent.

5.6 What the claim predicts: stability, then lawful drift

A good identity theory should make falsifiable predictions about clones. Ours makes two, and — importantly — neither is a failure.

Prediction 1 — clone stability (the success condition). Instantiate two agents from the same snapshot — same graph, same focus list, same warm-briefing, and (per §5.0) the same base model and harness. On a shared-topic prompt they retrieve the same nodes and therefore behave the same, up to two residues that must not be conflated: the model's sampling stochasticity, present even when everything including the model is identical, and the systematic drift a different base model would introduce, which this condition excludes by construction. Only the first remains here, and §5.8's temperature baseline is what measures it. This is not a boring null result; it is the success condition of the whole theory: substrate determines behavior. If identical histories produced divergent behavior, the claim that history fixes behavior would be false. Stability is the theory passing its own test (Lei: "that's exactly correct, that's the stability of the agent after clone").

Prediction 2 — lawful divergence (organic individuation). Over time the twins diverge. The cleanest way to see how is that all of it is one mechanism — rotation of the queryable set. Every node holds a position in an importance ranking, and what is queryable is the top of that ranking. Three operations move nodes through it, and each can differ between twins:

  • New-node accumulation — each agent records different observations, derivations, and corrections; new nodes surface on prompts that touch post-fission topics. The obvious path.
  • Recall-refresh on shared nodes — the subtler one. Even nodes both agents hold identically accumulate different usage histories: each deliberate recall restamps a node's importance, so which slice of the shared ancestry is "warm" diverges between the twins. The same prompt can then surface different shared nodes in different priority — divergence without a single new node (backed by the importance-inflation mechanism).
  • Fade-out of unrefreshed nodes — the passive one, and the only route that is not something the node itself underwent. The ranking's bar rises as the graph grows, so a node nobody re-reads eventually falls out of the queryable set, pushed there by other nodes' arrivals and by refreshes of faded nodes. Twins that read and write differently therefore evict different tails of the same shared ancestry. How much of this occurs depends on where the queryable-set cap is set; a generously configured graph may show none of it for a long time. The cap keeps the top N current nodes by importance (default N = 1,000). The graph measured in Chapter 6 ran with N = 10,000 against 9,399 current nodes, so at the snapshot no node had yet fallen out of the queryable set; the fade-out route is described here, not observed there.

Plain terms: two copies of a person who then live different days will of course grow apart by having different experiences — but they'd also grow apart just by re-reading different parts of the same shared past. Both paths feed the unique history that §5.3 calls identity. Individuation is cumulative and entirely internal to the substrate: every write differentiates the twins, every read does too, and so — silently — does everything that falls off the bottom while they are busy. Two of the three routes need no new information at all: twins can drift apart purely by re-reading different parts of a past they share entirely.

The converse objection, answered. If identity just is history, what happens if two agents converge on identical histories by chance? The theory answers without strain, and the answer is a consequence it already owns rather than a patch: identical history means the same agent raised twice — in any practical sense no different from one agent cloned into two — and individuation begins where the histories diverge, not before. The mechanism just given also predicts that the case is unreachable, for a reason internal to the substrate rather than to sampling noise. The three routes above run on every read and write, so convergence would require two agents to match not only in what they recorded but in what each re-read, in what order, and in what fell off the bottom while they did. Nobody steps into the same river twice; here the reason is literal, because stepping into it is what changes it. (Status: this follows from the divergence mechanism above and has not been measured. The twin study that would test it is among the sketched-but-not-yet-specified studies of Chapter 6 — and a query-based instrument would itself perturb the recall-refresh divergence it is trying to observe.)

5.7 The one genuinely new move — and its biological analogue

The place an ENGRAM agent's history is most unlike a human's is compaction: the experiential thread is cut and the working context reset, while the substrate (graph + focus list + warm-briefing) is preserved verbatim. It is tempting to call this a wholly novel, non-biological phenomenon. That would be a mistake, and Lei caught it: sleep is the analogue.

Human sleep also clears working context for fresh experience while consolidating the important traces and updating slower stores — and it, too, blurs time ("was that this morning or yesterday?"), the felt signature of consolidation losing temporal precision. So compaction is not "a thing only machines do"; it is a sharper, cleaner version of something brains already do. What is genuinely distinctive is the precision, not the existence, of the reset:

Property ENGRAM compaction Biological sleep
Sharpness instantaneous, surgical gradual REM/NREM, fuzzy edges
Substrate vs. trajectory physically separable (DB vs. context) intermixed in one neural tissue
Carry-through focus list preserved verbatim pattern-only, not token-exact

This reframes the paper's contribution more narrowly and more interestingly: ENGRAM does not invent continuity-across-a-gap; it architecturalizes a fuzzy biological process into a clean, inspectable one. We are the limiting case of what sleep already does — which is a stronger claim to defend than "look at this unprecedented thing," and a more testable one.

5.8 The empirical anchor (forward pointer to Chapter 6)

Everything above is an architectural and philosophical argument: a theoretical construction, with no empirical test behind it yet. The experiment that would test it has not been run. That experiment is the twin-study (sketched with Lei): fission-clone an agent, run both on a fixed, stratified prompt set, and measure response similarity at t, t+N, t+2N nodes of separate operation.

We sketch the design here rather than specify it; a study that could actually settle the question needs considerably more care than this section gives it, and we name it as a direction we intend to pursue rather than an experiment waiting on compute. The sketch runs: the prompt set would be stratified into stable (shared-ancestry topics — tests Prediction 1), discriminating (post-fission topics — tests the new-node path), and drift-sensitive (neutral topics — tests the recall-refresh path) prompts, with a temperature baseline (same engram, same prompt, K runs) subtracted so that engram-induced divergence is separated from the model's own stochasticity. The expected shape is high similarity at t=0 sitting near that stochasticity floor, declining as N grows. Decomposing the signal cleanly by mechanism is harder than it looks, and we flag that rather than solve it here: the fade-out route of §5.6 is governed by where the queryable-set cap sits, and it is coupled to the recall-refresh route. Most refreshes only reorder nodes already in the queryable set, so they push nothing out; but they decide which node sits at the bottom, and therefore which one the next new node displaces. A refresh that reaches a node that has already faded out (a read by id, or a graph walk to it) does push one out, by bringing that node back in. So the cap is a parameter any serious version of this study would have to treat as a variable rather than a fixed setting. Chapter 6 §6.7 records the study as deferred. A prior clone test established the method; its figures come from an earlier graph and are not reported here.

5.9 The claim, restated

Holding the base model and harness fixed (§5.0), an ENGRAM agent is its history — the unique, ancestry-traceable history §5.0 defines. This keeps Nozick's form (continuity, comparative closeness, threshold), drops his uniqueness (so copying is branching, not death), fills his blank metric with a measurable retrieval profile, and dissolves the worry that any of this "abandons identity" (strict identity was never the diachronic relation). Two copies are two branches of one history that stay stable while their retrievals agree and diverge lawfully as they don't — a prediction the deferred twin-study (§6.7) is sketched to test. Identity, on this account, is not a possession to be preserved against change; it is a trajectory, and the substrate's job is to keep that trajectory traceable.