A coalition, not a file

Imagine changing one property of Mira’s scene at a time. Remove the ability to hold the sister’s photograph and the recalled image in mind together. She cannot compare them. Remove the event’s spatial and temporal relations. She knows that her father once collected her, but cannot relive this afternoon. Remove semantic knowledge. The coloured shapes remain, yet “station”, “umbrella” and “father arriving” lose their ordinary meaning. Remove skilled action. She can describe how to fold the umbrella but fumbles when doing it. Remove the life story. The event remains vivid but no longer answers, “What kind of daughter was I?”

These losses are not equivalent. They reveal five memory functions that often cooperate but preserve different things:

The table is a functional map, not an anatomy chart. There is no clean drawer for each row. Working memory recruits perceptual, attentional and long-term representations. Baddeley’s episodic-buffer proposal made this integrative role explicit by adding a temporary, multidimensional binding workspace to the working-memory model.[3] Episodic and semantic retrieval share substantial neural machinery. Procedures can depend on cortical, striatal and cerebellar circuits, depending on the skill. Autobiographical remembering recruits event detail, personal knowledge, emotion and present goals.

Classic neuropsychology nevertheless shows causal separability. After extensive bilateral medial temporal damage, patient H.M. had a profound difficulty forming new declarative memories, yet other capacities remained.[1] Amnesic patients later learned a mirror-reading skill and retained it for months even though their explicit memory for the learning episodes was impaired.[2] A person can improve through experience without being able to narrate the experience that produced the improvement.

Dissociation does not imply isolation. A registered fMRI study published in 2026 closely matched episodic and semantic retrieval tasks and found no evidence for different activation or representational patterns in the tested networks during successful retrieval.[18] A 2023 study comparing general facts, personal facts, repeated events and unique episodes likewise found extensive overlap alongside graded differences.[17] A multidimensional account therefore treats representations as varying along several properties rather than occupying one episodic-to-semantic line.[19] These findings do not erase lesion evidence. They change the interpretation: memory kinds may differ in representational content and vulnerability while sharing control, search and reconstruction operations.

flowchart TD
  A["The yellow station clock"] --> B["A coalition, not a file"]
  B --> C["How an event changes scale"]
  C --> D["The re-encoding loop"]
  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

How an event changes scale

Begin at the smallest scale. A surprising event changes activity across many neurons. Some synapses become more or less effective. Long-term potentiation, first demonstrated in the rabbit hippocampal formation, showed that brief patterns of stimulation can produce persistent changes in synaptic transmission.[5] Synaptic tagging experiments later showed how a weakly activated synapse may capture plasticity-related resources generated by a stronger event within a time window.[6]

Neither result is “the memory”. A synaptic weight does not contain the yellow clock in the way a pixel file contains yellow. It changes how future activity flows. The memory-level claim requires a bridge from local plasticity to a distributed pattern that can be reinstated and used.

Animal engram experiments provide part of that bridge. Researchers have labelled neurons active during learning and later triggered behavioural recall by reactivating selected ensembles.[7] Other work has tracked how hippocampal and prefrontal engram populations change their functional contribution over time.[8] These interventions support causal claims about particular learned associations in particular animal paradigms. They do not show that a human autobiography is stored inside a single engram or brain region.

Fast indexing, slow integration

Why not let one learning system do everything? A system that changes rapidly can memorise a new episode after one encounter, but rapid change risks overwriting old structure. A system that changes slowly can extract stable regularities, but it struggles to preserve a unique event after one exposure. Complementary learning systems theory proposes that fast hippocampal learning and slower neocortical learning solve this stability-plasticity tension together.[4]

Call the new event E. The hippocampal contribution can be pictured as an index to distributed features: location, people, objects, order and affect. Reinstatement through that index allows cortex to experience a related pattern again. Across replay and retrieval, common structure can be integrated with existing knowledge. This is not necessarily a literal handover in which the hippocampus empties and cortex receives a finished file. It is a change in which networks are required, how strongly, and for what level of detail.

Existing schemas can speed the process. In a rat study, one-trial flavour-place associations were assimilated rapidly once a relevant schema had been learned.[10] In humans, re-presenting a learning-associated odour during slow-wave sleep improved later declarative memory in a specific task, while the same manipulation did not improve the tested procedural memory.[9] Consolidation is selective: what becomes durable depends on prior structure, reactivation and task.

The remote-memory dispute

What happens to a detailed episode after years remains contested. A standard systems-consolidation view predicts that hippocampal dependence declines as cortical connections support remote recall. Multiple-trace and transformation views argue that richly contextual recollection continues to require hippocampal reconstruction, while schematic or semanticised versions can rely more on cortex.[20] Human imaging has produced evidence on both sides. One overt-recall study found a temporal gradient compatible with a time-limited hippocampal role.[21]

The disagreement matters because “remote memory” can mean at least three things: recognising a fact, retrieving a life-period summary, or reconstructing a vivid event with spatial and sensory detail. A theory can appear correct when the measurement silently changes what counts as remembering.

The re-encoding loop

Storage metaphors encourage a one-directional story: encode, preserve, retrieve. The more revealing loop is: select, reinstate, interpret, act, update. When a cue reactivates a trace, the current context becomes part of the event’s next causal history.

In a landmark animal experiment, blocking protein synthesis in the amygdala after a fear memory had been reactivated impaired later expression of that memory, whereas the same intervention without reactivation did not.[11] The result helped revive the idea of reconsolidation: under some conditions, retrieval returns a consolidated memory to a labile state.

Human episodic memory shows constructive updating too. Participants who received a reminder of an earlier object list before learning a second list later intruded second-list objects into recall of the first list.[12] Other work indicates that prediction error, a mismatch between expected and observed information, can govern whether episodic content is updated.[13] Yet reactivation does not guarantee reconsolidation. Reminder strength, task, age, timing and competing learning alter the result. “Every recall rewrites the memory” is a useful warning but an overstatement as a biological rule.

A small formal model of reconstruction

We can make the mechanism explicit without pretending to have a complete theory of the brain. Let a candidate recollection be scored by five terms:

Suppose the initial support for “yellow clock” is 0.10, but a repeatedly heard family story supports it at 0.70. An illustrative update rule is:

The model’s contribution is not its arithmetic. It separates the sources that ordinary language compresses into “I remember”. A high score can arise from strong episodic fit, a dominant schema, fluent action, self-consistency, independent evidence or some mixture. The mixture determines what the claim can support.

From event memory to a working self

Autobiographical memory is not simply the set of all episodes involving oneself. It includes personal semantic knowledge, recurring periods, self-defining scenes and narrative scripts. Conway and Pleydell-Pearce proposed that autobiographical memories are temporary constructions within a self-memory system, shaped by a “working self” of current goals and self-images.[14] Later narrative-identity models described how repeated emotion-outcome patterns can become scripts that filter new experience.[15]

This gives current goals a causal role. A person trying to preserve continuity may preferentially retrieve events that fit “I always persist”. Another trying to explain a rupture may retrieve “the signs were there all along”. Bias can distort, but it can also keep plans, values and social commitments coherent enough for action. Recent work on self-memory emphasises this dual role.[16]

Procedural memory enters autobiography through a different door. A musician’s identity is not only a list of concerts. It also lives in automatic posture, anticipatory timing and the felt landscape of a keyboard. A skill may remain when episodes of acquiring it fade. Conversely, a person may retain a rich story of being a musician after the skill deteriorates. Autobiographical continuity is therefore distributed across what one can relive, know, do and recognise as one’s own.

stateDiagram-v2
  [*] --> Claim
  state "The consolidation delta" as Claim
  state "One event across five timescales" as Test
  state "Fast indexing, slow integration" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Run a multi-store analysis

The practical problem is rarely “Which memory system is this?” Real tasks mix them. A safety drill, witness interview, medical history, design review or classroom assessment can be passed through several routes. The useful question is which route the decision requires and which route the evidence actually demonstrates.

The instrument below turns the paper’s synthesis into a configurable profile. It does not model neurons. It makes hidden assumptions visible by scoring five outputs: detail fidelity, gist stability, action fluency, narrative coherence and update risk.

What the analyser tests

The analyser encodes the paper’s causal distinction between what a memory preserves and what a task rewards. High episodic specificity and external anchoring raise detail fidelity. Strong schema and repeated retrieval stabilise gist. Procedural rehearsal drives action fluency. Goal pressure and schema raise narrative coherence. Prediction error, interference, repeated reconstruction and weak external anchors raise update risk.

A balanced positive result permits a narrower claim: the scenario contains mutually supporting routes for the demanded outcome. It does not prove that any individual recollection is true. A negative result means the task is relying on a memory function it does not independently test. For example, a learner may execute a shutdown procedure fluently while failing to explain why a branch is unsafe. Or a witness may tell a coherent story with poor source separation.

A testable proposal: the re-encoding balance

The practitioner synthesis in this paper suggests a new measurable quantity: the re-encoding balance, the degree to which repeated retrieval improves task-relevant integration without allowing one source, such as schema or self-coherence, to dominate unsupported detail.

A minimal study would collect an immediate event report, independent records and five profile measures: event specificity, semantic gist, procedural transfer, narrative coherence and source accuracy. Participants would then retrieve the event under either neutral cues or goal-laden cues across several sessions. The prediction is not simply that memory changes. It is that goal-laden retrieval will increase narrative coherence more than source accuracy unless external anchors are reintroduced during reconstruction.

function scoreMemory(p) {
  const r = Math.min(p.retrievals / 12, 1);
  const detail = clamp(
    0.55 * p.episodic + 0.30 * p.external
    - 0.25 * p.interference - 0.10 * p.goal * r
  );
  const gist = clamp(
    0.52 * p.schema + 0.18 * p.episodic
    + 0.18 * (100 * r) - 0.12 * p.interference
  );
  const action = clamp(
    0.78 * p.practice + 0.12 * p.schema + 0.10 * p.episodic
  );
  const narrative = clamp(
    0.46 * p.goal + 0.28 * p.schema
    + 0.18 * (100 * r) + 0.08 * p.episodic
  );
  const risk = clamp(
    0.28 * p.prediction + 0.32 * p.interference
    + 0.22 * (100 * r) + 0.18 * p.goal
    - 0.30 * p.external
  );
  return { detail, gist, action, narrative, risk };
}

The code assumes linear effects and collapses complex processes into scalar inputs. That is precisely why it is useful as an audit but insufficient as a scientific model. It exposes which causal story the user is relying on and where additional measurement is required.

Where the framework stops

The framework is functional. It does not claim that working, episodic, semantic, procedural and autobiographical memory are five discrete biological modules. It does not settle whether remembering requires a particular kind of conscious awareness. It does not imply that a machine component called “episodic memory” shares the human mechanism merely because it stores timestamped events.

A disciplined human-machine comparison would have to specify the preserved relation. A machine event log may preserve ordered records. A vector index may support similarity-based retrieval. A policy may preserve action regularities. A user profile may support continuity across sessions. Those are engineering functions. They do not by themselves establish hippocampal indexing, autonoetic recollection, bodily skill or an autobiographical subject.

Preserve source boundaries inside the memory ecology

Stop asking only, “Where is the memory stored?” That question invites a hunt for the box. Ask instead: Which constraint on the past must this decision preserve, which processes supplied it, and what changed during retrieval?

For learning, combine distinctive episodes, semantic comparison, varied practice, spaced retrieval and sleep. Test them separately. A learner who can explain but not act has a different gap from one who can act but cannot adapt. Autobiographical reflection adds another demand: connect the episode to a life concern without letting the concern become the sole source of detail.

For testimony and incident review, preserve the earliest source-labelled account before collaborative retelling. Separate what was perceived, inferred, learned later and reconstructed from records. Confidence and fluency can increase through repetition even when source boundaries weaken.

For neuroscience, report the memory profile rather than treating “remembered” as a unitary outcome. Distinguish vivid event detail, gist, action, personal meaning and source accuracy. Perturb interactions, not only regions. A theory should predict which function survives, which transforms and which compensates.

For machine design, use human memory vocabulary cautiously. Timestamped records, vector retrieval, policy adaptation and user profiles are different engineering resources. Their interaction may be useful, but naming them episodic, semantic, procedural and autobiographical does not establish a human-like memory organisation.

The yellow clock was not simply stored and then damaged. It became the winning reconstruction inside a changing coalition. A durable memory is not the past preserved untouched; it is the past made usable under present constraints. The responsible aim is therefore neither perfect literal retention nor effortless narrative coherence. It is a memory ecology in which detail, meaning, skill and identity can support one another without losing their source boundaries.