When the same moment is not the same state

A snapshot becomes misleading when it records what is present but omits how the present acquired its dispositions.

The hidden assumption inside a snapshot

Imagine a thermometer showing 20°C. Is the object warming or cooling? The display does not say. Add a reading from one minute earlier and the direction becomes visible. Add the control setting and you may predict the next minute. Nothing supernatural appeared. The original state description was simply too thin for the decision.

The same problem recurs in cognition. A word can be a promise, a threat, a quotation or a joke. A facial expression can continue an emotion or mark its reversal. A model token can repeat an earlier proposition or retract it. At the level of the current signal, these cases may match. At the level of an ongoing process, they differ.

Bergson’s first intervention is aimed at the picture that made the reset listener seem normal. Clock time places instants beside one another in a homogeneous line. Lived time does not arrive that way. In his account of duration, phases of consciousness differ while permeating one another. Separating them as if they were beads changes the phenomenon being described.¹³

Duration is more than a longer memory buffer. A buffer can hold independently addressable items. Bergson’s stronger claim is that earlier phases survive by modifying the qualitative whole in which the next phase occurs. The final note does not sit beside four remembered notes. The phrase has become a directed musical experience.

Duration and occasions: two corrections to the thing-model

Bergson begins from lived continuity. Whitehead begins from becoming and relation. Their overlap is useful only when their disagreement remains visible.

Bergson: a past that changes the quality of the whole

Count five sheep and each can be placed beside the others. Count five seconds and the notation invites the same treatment. Bergson argues that this spatial picture works for measurement but misdescribes immediate experience. A felt interval is not assembled from self-contained temporal atoms. It develops as a qualitative multiplicity in which phases are heterogeneous, continuous and irreversible.¹³

Consider anger that slowly becomes embarrassment. The later feeling contains the earlier orientation but transforms it. If we isolate “anger at 10:02” and “embarrassment at 10:03” as independent objects, we can classify them, but we lose the transition through which one became the other. The transition is not decorative glue. It explains the meaning of the later state.

This is why Bergson repeatedly uses movement and melody. An analysed sequence can be represented as symbols, yet the symbols are partial expressions of a flow whose moments are not repeatable units. His proposal is phenomenological and metaphysical. It starts from the structure of lived change, then challenges a world-picture built from static elements.²

Whitehead: an event that inherits a world

Whitehead makes a different move. Instead of asking how stable states flow, he treats becoming as primary and stable things as organised patterns of events. In his mature ontology, an actual occasion is an elementary process of taking account of an antecedent world and integrating many relations into one determinate event. He calls this growing-together concrescence.⁵⁶

A chair, organism or person is therefore not one enduring lump. It is a structured continuity, a “society” in Whitehead’s technical vocabulary, reproduced across occasions. Persistence is an achievement of patterned inheritance. The past is not retrieved from a warehouse after an event has formed. It is among the relations from which the event becomes what it is.

Whitehead’s account is not simply Bergson with more nouns. Bergson resists decomposition of duration into juxtaposed units. Whitehead’s mature view gives becoming an atomic structure. Bergson begins from the continuity of lived experience. Whitehead builds a systematic cosmology in which physical and mental nature belong to one process scheme. Whitehead also advances a panexperientialist metaphysics that ordinary cognitive evidence does not establish.⁶

flowchart TD
  A["When the same moment is not<br/>the same state"] --> B["Duration and occasions: two<br/>corrections to the thing-model"]
  B --> C["The process-sufficiency test"]
  C --> D["Keeping the right past"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#B96D00,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The process-sufficiency test

The practical question is not “Is everything a process?” It is “Does this decision remain invariant when relevant history is changed under a matched present-state description?”

From intuition to a testable distinction

Two phrases are often blurred. Chronological storage means that earlier observations remain available somewhere. Historical constitution means that earlier events have changed the present causal organisation or relation through which the next event forms. A log has chronological storage. A trained habit, a changed synapse, a live contractual obligation and a recurrent latent state can exhibit historical constitution.

The distinction matters because adding a longer transcript does not guarantee process sensitivity. A model can receive every earlier token yet ignore their order. Conversely, a compact recurrent state can carry the effect of a long history without retaining its episodes. Sequence length is not the same property as historical constitution.

If ΔH exceeds a decision-relevant tolerance and repeats under controls, the measurement M is insufficient. Three repairs are possible. Expand the current state, preserve a trajectory, or enlarge the system boundary to include the missing relation. A process model is not automatically a model with no states. It is often a model whose states are explicitly understood as condensations of prior activity.

A credible history intervention

The formula is compact, but an experiment can fail in several ways. Suppose History A contains praise and History B contains abuse. If the systems differ later, the cause may be emotional learning, lexical frequency, safety policy or a hidden tool result. Calling the difference “duration” does not separate these rivals. The intervention must specify what changed, what was matched and which future outcome carries the claim.

Begin with a candidate boundary. For a human listener, it might include current stimulus, recent neural state and task context. For an artificial agent, it might include model version, runtime state, retrieved evidence, authoritative world state and available tools. The boundary is not chosen by convenience alone. It must contain every variable that could plausibly mediate the outcome, or explicitly mark what remains outside.

Next choose two histories that differ in one causal property. Reorder the same cues to change temporal structure while preserving cue counts. Insert a meaningful event boundary without changing the sensory item at the probe. Reset the recurrent state while retaining the transcript. Remove a signed commitment while keeping the language that referred to it. These interventions expose different kinds of historical constitution.

Then match the declared present. Natural matching finds cases that already share the same measured state. State cloning copies a defined interface between systems. A surgical intervention sets the suspected carrier directly. Each method has a different weakness. Natural matches can hide confounders. Cloning can omit latent variables. Surgical changes can create states the system would never reach on its own.

Finally, use controls that can defeat the preferred explanation. Repeat identical histories as a negative control. Swap an irrelevant earlier segment to test generic sensitivity to disturbance. Include a rival generator that has chronological access but no order-sensitive update. Probe more than one future outcome, because a history may affect verbal report while leaving action unchanged. A process claim earns force when the effect follows the altered relation, disappears under the negative control and survives a plausible rival.

What current science can and cannot add

Cognitive science supplies mechanisms that make the general problem concrete, not a verdict on either philosopher. Research on temporal receptive windows shows that different cortical regions integrate information over different durations during narrative comprehension. Short-window regions track local features, while longer-window regions depend on information accumulated across sentences and scenes.⁸

Event-segmentation research adds a complementary operation. Continuous activity is parsed around prediction errors, contextual shifts and changes in goals. Event boundaries can reorganise memory rather than merely mark an external timestamp.⁹¹⁰ In four experiments, Pu and colleagues found that boundaries improved temporal-order memory within events and impaired it across events. A temporal-context model captured the pattern by allowing context to drift and reset at boundaries.¹²

A more recent registered report examined perceptual continuity and event segmentation in the same paradigm. Conceptual boundaries reduced serial dependence even without sensory change, but individual differences in continuity and segmentation were not reliably coupled. The authors treated the null relation as inconclusive because the measures themselves had low reliability.¹³ That result is a useful brake on an attractive story: continuity and segmentation may interact without being one mechanism.

Keeping the right past

A useful process architecture does not remember everything. It preserves the causal residues that remain valid, authorised and decision-relevant.

A configured-system example

Consider a support assistant handling the sentence “That’s fine.” In one conversation, the customer has calmly accepted a revised delivery date. In another, the customer has repeated a complaint three times and uses the same words after being ignored. A turn-level classifier may assign the same sentiment and intent. A process-sensitive system represents the event trajectory: unresolved issue, failed repair attempt, rising interaction risk and current authority to offer a remedy.

The correct design is not to dump the entire transcript into a model and call it memory. The system needs distinct carriers. The transcript is evidence. The event model groups turns into episodes. The history-bearing state tracks unresolved commitments and interaction phase. Authoritative world state records whether a refund, reshipment or complaint case actually exists. A policy layer determines what action is permitted. Receipts record what changed.

Executable lab: history, decay and reset

The lab below exposes the logic hidden by the phrase “context matters”. It runs two opposed histories that end on the same current cue. The decay control changes how much prior input survives. The boundary control erases a proportion of accumulated context just before the current cue. The history preset changes order while preserving the current cue.

from math import tanh
from typing import Iterable, Tuple


def run_history(
    values: Iterable[float],
    rho: float = 0.6,
    reset: float = 0.0,
    gain: float = 2.0,
) -> Tuple[float, float]:
    """Return final latent state and bounded outcome.

    Invariant: 0 <= rho < 1 and 0 <= reset <= 1.
    The reset is applied immediately before the final cue.
    """
    xs = list(values)
    if not xs:
        raise ValueError("history must contain at least one cue")
    if not 0 <= rho < 1:
        raise ValueError("rho must be in [0, 1)")
    if not 0 <= reset <= 1:
        raise ValueError("reset must be in [0, 1]")

    h = 0.0
    for index, x in enumerate(xs):
        if index == len(xs) - 1:
            h *= 1.0 - reset
        h = rho * h + float(x)
    return h, tanh(gain * h)


def process_sensitivity(a, b, **kwargs) -> float:
    _, y_a = run_history(a, **kwargs)
    _, y_b = run_history(b, **kwargs)
    return abs(y_a - y_b)


A = [1, 1, -1, -1, 0]
B = [-1, -1, 1, 1, 0]
print(run_history(A))
print(run_history(B))
print(process_sensitivity(A, B))

The failure case: a past that should have ended

History dependence is not automatically intelligence. A support assistant may carry anger from a resolved complaint into an unrelated purchase. A clinical model may let a superseded note dominate current measurements. An agent may continue a plan after its authority expired. In each case, persistence converts yesterday’s relevance into today’s error.

Process architecture therefore needs segmentation as much as continuity. A boundary policy decides which state decays, which state resets, which evidence remains retrievable and which obligation persists. A process without boundaries becomes contamination. A snapshot without inheritance becomes amnesia.

A decision surface for process claims

Do not promote every sequential effect into a theory of process. Evaluate both the size of the history effect and the repeatability of the intervention. The matrix below separates a real process signal from noise, confounding and unnecessary architectural complexity.