To a seven-year-old
Knows: that people remember and forget things, and that a smell or a song can suddenly remind you of something — but not that memory has parts, or that a memory changes after you make it.
I want to tell you about a party.
There was a song playing. My friend was there, and we talked about a big question for a long time. The room was crowded. The kitchen smelled of something cooking. It got very late — nearly two in the morning. And the whole thing felt like a good night out.
That's not one memory. That's seven little pieces, and I keep them in a jar.
Look at the panel. Every bubble is one piece. The threads between them are glue. The glue is what makes them one memory instead of seven separate ones.
Tap the song
Go on — tap the orange bubble, the song. Watch what happens. The song lights up first, then the glue carries the light along the threads to my friend's face, then to the talk, then out to the room and the smell. One tug, and the rest come along.
That's what a reminder is. You only need to grab one piece. The glue brings you the others for free. That's why a song on the radio can hand you a whole night from years ago.
Now try losing one
Drag the smell of the kitchen away from the middle. Its thread goes thin, thin, thin — and the bubble fades out. Now tap the song again. The smell doesn't come back. It slipped off.
That's forgetting. Not a hole where the memory was. Just a piece that wasn't glued on tightly enough to be tugged.
The glue button
Here's the good bit. Press think harder about it. Every thread gets fatter. Now try dragging a bubble away — it fights you, and it comes back.
The glue gets stronger when you really think about what a thing means. Not when you just say it over and over. Thinking about the meaning is thick glue. Saying a word again and again is thin glue.
Tap around. Drag things off. Then press the button and try again. Watch the number under the jar: how many of my seven pieces did you get back?
To a teenager
Knows: cramming works badly and testing yourself feels hard — but not the vocabulary of encoding, storage and retrieval, nor that shallow repetition is a real, named failure mode.
Last level I said a memory is pieces held together with glue, and that a reminder tugs one piece and the rest come with it. Those were stand-ins. Here are the real names.
Four phases, in order
Encoding is the moment the party goes in — what you actually did with the song and the conversation while they were happening. Consolidation is the slow settling afterwards, over hours and then years, when you are not thinking about it at all. Retrieval is getting it back out. And forgetting is what happens when one of the other three doesn't deliver.
That last sentence is the whole point of this level. "I forgot" is not one thing. It is a result, and there are several different upstream causes, and the study advice you need depends entirely on which one bit you.
"Gluing harder" was depth of processing
Craik and Lockhart, in 1972, proposed that how well you remember something is a function of how deeply you processed it 1. They named three levels. Structural: what it looks like — the shape of the word philosophy, the capital letters, the length. Phonemic: what it sounds like — that it rhymes with nothing much and has four syllables. Semantic: what it means — that your friend was using it at 2am to argue something you half agreed with 1.
Deeper processing produces a stronger, more elaborate memory 1. In the panel, pick the route yourself: send philosophy through structural, phonemic or semantic encoding and watch the trace bar that comes out the other side.
Why re-reading feels like studying and isn't
Saying a word to yourself over and over has a name: maintenance rehearsal. Craik and Lockhart argued that rehearsal which just repeats previous analyses does not enhance long-term memory 1. Re-reading your notes is maintenance rehearsal wearing a costume. Time spent with the material is not the currency; what you did with it is.
The cue has to match
Now the part nobody tells you. A deep trace is not enough on its own, because retrieval depends on the cue — and cues only work if they match how you encoded. That is the encoding specificity principle (Tulving & Thomson, 1973): when people encode information, they do so in specific ways 2. Encode by sound, get tested by meaning, and you can fail a test on material that is sitting right there in your head.
So in the panel, after you choose the encoding route, choose the cue at test: nothing, a rhyme, a meaning, or the original context — the crowded room, the kitchen smell, 2am. Then press run test. The phase strip lights up whichever phase actually failed.
Two flavours of forgetting
Psychologists split forgetting into two families: either the memory has disappeared and is no longer available, or it is still stored but for some reason cannot be retrieved — not accessible 3. Never encoded deeply? Availability problem; no cue on earth will save you. Encoded deeply but cued wrongly? Accessibility problem; the item is intact and you just took the wrong road to it.
Run the panel four or five times. The verdict line will tell you which failure you engineered — and that is the difference between "study differently" and "practise retrieving".
philosophy, lifted straight out of the party conversation.To a cognitive psychology undergraduate
Knows: the multi-store model and basic experimental design — but tends to treat 'depth' as self-evident and forgetting as a single phenomenon.
Level 2 told you that thinking about meaning glues a memory harder. That sentence is doing a lot of unpaid work. Let's make it an experiment.
Depth is a manipulation, not a metaphor
Craik and Lockhart's 1972 levels-of-processing model "describes memory recall of stimuli as a function of the depth of mental processing, where deeper levels of processing produce more elaborate and stronger memory than more shallow levels of processing" 1. The operational trick is that you never ask the participant to remember. You ask an orienting question, and the question fixes the level: structural (how is the word spelled, how do the letters look), phonemic (does it rhyme with fall), semantic (does it fit this sentence) 1. Recall is then incidental, and the only thing you varied was the analysis performed at study.
The claimed consequence is about the trace, not the effort: "Shallow processing (e.g., processing based on phonemic and orthographic components) leads to a fragile memory trace that is susceptible to rapid decay. Conversely, deep processing (e.g., semantic processing) results in a more durable memory trace" 1. That is why the model was a challenge to time-in-store accounts: Craik and Lockhart "argued that rehearsal that consists simply of repeating previous analyses (maintenance rehearsal) does not enhance long-term memory" 1. Press the maintenance rehearsal only toggle in the panel: every element gets rehearsed, dwell time goes up, durability does not.
Retrieval is conditional, not a strength readout
Now the second half, the half undergraduates skip. Even a durable trace can be silent. The organising principle is encoding specificity: "when people encode information, they do so in specific ways" (Tulving & Thomson, 1973) 2, and cues work to the extent they match those specifics. The party is the textbook's own example — "perhaps you heard it while you were at a terrific party, having a great, philosophical conversation with a friend. Thus, the song became part of that whole complex experience" 2 — which is why a song on the radio can "suddenly evoke memories of an earlier time in your life, even if you were not trying to remember it when the song came on" 2.
So model recall as a product, not a sum:
# per element i, given cue set C d_i = durability(level_i) # .20 struct, .45 phon, .85 sem 1 a_i = d_i * exp(-t / (30*d_i)) # AVAILABILITY after interval t m_i = overlap(C, context_i) # ACCESSIBILITY: encode-cue match 2 P(recall_i) = a_i * (0.15 + 0.85*m_i)
Two factors, two failure modes, and the panel labels each miss with the one that killed it. a_i low = never encoded deeply (unavailable). a_i high but m_i low = encoded but not cued (inaccessible). This is the availability/accessibility dichotomy, and it is not a philosopher's distinction: "Something that cannot be retrieved now and which is seemingly gone from memory may, with different cues applied, reemerge" 2. Set every element to semantic, run the interval out to a week, then cycle the cue set from none to full. The traces never changed. The total triples.
Three kinds of forgetting, unequal support
Decay "states that forgetting occurs as a result of the automatic decay or fading of the memory trace" 3 — that is the exponential in the code above, and you should distrust it: "There is very little direct support for decay theory as an explanation for forgetting in short-term and long-term memory. It is also hard to test. In practice, researchers cannot create a truly blank period between presenting material and testing recall. Once information is presented, participants rehearse it" 3. You cannot buy a blank interval. Displacement is a capacity story instead: when short-term memory is full, "new information displaces or 'pushes out' old information and takes its place" 3. And cue-dependent failure is the one this panel makes you feel 2.
To a PhD student in cognitive neuroscience
Knows: LTP, hippocampal anatomy, retrograde amnesia gradients — but may still treat the two-process consolidation split as settled and the gradient as a clean measurement.
Level 3 left "consolidation" as a single box in a phase strip. That box hides two processes that share almost nothing except a name — they run on different timescales and at different levels of organisation, and the evidence for each comes from different experiments.
Process one: hours
The fast one is synaptic (cellular) consolidation: "a cascade of cellular and synaptic neurochemical events initiated by learning and ending within hours at most" 4. Mechanistically the standard account is that learning activates intracellular signalling that changes gene expression and protein synthesis, producing "lasting alteration of synaptic proteins, as well as synaptic remodeling and growth" 5. The plasticity currencies are the ones you already know: long-term potentiation strengthens synaptic connections, long-term depression weakens them, and these changes "are thought to underlie the initial formation of memory at the cellular level" 5.
The consequence you should carry into the panel is the fragility window. Immediately after the party, before the protein-synthesis-dependent alterations have run, the trace is disruptable 5. Minutes and hours, not years. Wikipedia's framing puts it at "the first few hours of learning" 5.
Process two: decades
The slow one is systems consolidation, "sustained over much longer periods of time, sometimes decades," involving "the reorganization of memory at the level of large-scale neural systems" 4; the same two-process split, at the coarser grain of hippocampus-to-cortex reorganisation over weeks to years 5.
Note that the standard model is actually two models wearing one coat. The hippocampus "plays a time-limited role in memory storage, either by first storing and then transferring a memory, or by acting as an index … that binds extrahippocampal storage sites during a period when direct communication among these sites is inadequate to permit memory retrieval" 6. Store-then-transfer and index-then-release make different commitments about what is in the hippocampus at t=0. Most gradient data do not distinguish them.
A third process, reconsolidation, has been proposed: "previously consolidated memories can become unstable again after retrieval and may need to be updated before being re-stabilized" 5. Treat that as a proposal in this explainer — the evidence I am citing states it, it does not demonstrate it. There are no reconsolidation primary data in my ledger, and I will not pretend otherwise.
The problem the panel is about
Here is the thing a supervisor says at the whiteboard. Everyone quotes "the consolidation period" as though it were a physical constant of the memory. It is not. It is a fitted parameter of an experiment. "The apparent extent of RA, hence the length of the consolidation period, varies with both the memory task being tested and the method used to produce amnesia, e.g., lesions, electroconvulsive shock, immersion in cold water, or the injection of protein synthesis inhibitors" 6.
So drive it. In the panel, pick an amnestic method, pick a probe — vivid episodic recall of the party versus gist recognition — set lesion timing, set cohort n, and press run 100 subjects. Watch the fitted gradient length settle with its interval. Then change only the method and watch the same underlying memory yield a different number; the log line will tell you how far it moved. Change only the probe and it moves again.
Two habits to take away. First, the four methods are not four ways of measuring one thing: a protein-synthesis inhibitor targets the hours-long synaptic cascade 45, a hippocampal lesion targets the decades-long systems process 4. They can return gradients differing by an order of magnitude because they are interrogating different processes. Second, gist survives lesions that flatten detail — flip the probe toggle and the fitted gradient shortens dramatically, which is exactly the observation level 5 will use to argue about transformation.
Flip reactivate first and the pre-manipulation retrieval reopens a vulnerability window on an otherwise old memory 5. The simulation implements that as a modelling assumption, drawn in dashed lines. Do not read it as data. And note the last caveat in the same source: the animal literature and human timescales do not line up cleanly 6.
To a peer — a memory researcher
Knows: the whole literature and its politics; wants to know which parts of the level-4 story are folklore, which results actually adjudicate, and where this evidence base runs out.
Let me retract something. At level four I drew consolidation as two clean processes on two clean timescales, and then made a virtue of admitting that the retrograde gradient wobbles. That framing is still too generous to the standard model. The two-process picture is a theoretical commitment that organises data; it is not itself a finding. When I run the party memory forward twenty-five years in the panel, what I am really doing is drawing one theory's curve and calling it the phenomenon.
What the standard model actually asserts
Stated properly it is a claim about links, not about a location: "over time, the links among the neocortical elements of the ensemble that constitutes the content of the memory trace, guided/reinforced by the HPC, are strengthened to the point that they can be reactivated without hippocampal input. This marks the end of the sy[stems consolidation period]" 4. Note the terminal clause. The model needs an end — a point where the hippocampus is no longer required. That is what makes it falsifiable, and that is exactly where it takes fire.
The model comes in two flavours, and peers routinely conflate them: the hippocampus "plays a time-limited role in memory storage, either by first storing and then transferring a memory, or by acting as an index ... that binds extrahippocampal storage sites during a period when direct communication among these sites is inadequate to permit memory retrieva[l]" 6. Store-then-transfer and index-then-release make different predictions about partial lesions and about what a degraded trace should look like. Treat them separately or your experiment adjudicates nothing.
The transformation alternative
The rival is not "the hippocampus stays involved forever" — that is a caricature people attack. It is a claim about content: "the hippocampal and neocortical systems are critical for different forms of memory, and that the shift of memory from dependence on hippocampus to dependence on neocortex during consolidation is a reflection of the fact that memory often is transformed with time, becoming more generic in nature" 6. So the observed time-by-region interaction is real; the interpretation is that your probe drifted, not that a trace migrated. Twenty years on you are no longer asking about the party. You are asking about a good night out.
This is the panel's whole point. Drag memory age and flip the probe. Under gist probing the two curves lie almost on top of each other for most of the range — press construct a discriminating case and watch the machine hunt for the narrow band where they separate. If your design lives outside that band, your null is uninformative, whatever your n.
Who is winning
I will not pretend neutrality. Contrasting the two approaches and evaluating "their fit to the data from a number of approaches," one review concludes "that the data from neuroimaging studies strongly support multiple trace theo[ry]" 6. I take that seriously and I also flag its status: the paper is a 2007 review 6, and nothing in this evidence base establishes how that assessment has held up since. Cite it as a position, not as a settled score.
The stronger move is to give up the construct. One recent proposal: "the concept be replaced by one of memory systems reorganization, which does not carry the theoretical baggage of systems consolidation and is flexible enough to capture the dynamic nature of memory from inception to very long-term retention and retrieval at a psychological and neural level" 4. Non-fixed traces, no terminal point, no gradient to measure. I find this attractive and I also notice the cost: a framework flexible enough to capture everything is harder to break.
Where this ledger runs out — press the button
Press show the ledger's silence and half the arena greys out. There are no engram or optogenetic records here. No sleep-dependent replay parameters — even though replay is invoked as the medium of the hippocampal–neocortical dialogue in the very passage that defines the standard model 6. No synaptic tagging-and-capture. No reconsolidation primary data, so my level-four "third process" was a proposal I dressed as a mechanism. And no quantitative time constants: the ten-year decay in the panel is an illustration, not an estimate.
What the ledger does contain is the methodological wound: "the apparent extent of RA, hence the length of the consolidation period, varies with both the memory task being tested and the method used to produce amnesia, e.g., lesions, electroconvulsive shock, immersion in cold water, or the injection of protein synthesis inhibitors," with the animal literature complicating matters further 6. Task-dependence is not noise. It is the transformation prediction, arriving as a confound in someone else's experiment.