The three compartments maintain distinct but coupled state variables across three time horizons.
At the fast scale:
- the presynapse tracks residual calcium and its readily-releasable vesicle pool — both of which encode the very recent history of firing.
- the postsynapse tracks membrane voltage and the amplitude and speed of its calcium rise, which together encode the instruction for future change.
- the astrocyte tracks glutamate concentration both inside and outside the cleft, its own local and global calcium state, and its fuel output.
At the intermediate scale, the shared signal layer tracks whether the mGluR overflow sensors have fired, and whether the neuromodulatory context gate has been set — specifically whether PKA has primed the AMPA insertion machinery, silenced the forgetting phosphatase, and enabled gene expression in the nucleus.
At the slow scale, all three compartments track their own physical architecture:
- the postsynapse its receptor count and spine size
- the presynapse its docking slot count and vesicle channel clustering
- the astrocyte its wall distance from the synapse, its matrix density, and its baseline co-agonist supply.
Every action potential sets off a precise sequence across all three compartments simultaneously.
The presynapse converts the electrical event into a chemical wavefront: calcium floods in, drives probabilistic vesicle release from the readily-releasable pool, and leaves a residual trace that biases the next release upward if spikes keep arriving and downward if they stop. The amount of glutamate released fills the cleft and begins diffusing outward.
The astrocyte responds in two parallel arms the moment glutamate spills beyond the cleft boundary. The first arm activates astrocytic mGluR5 receptors via a Gq cascade, triggering an internal calcium rise that is directly proportional to how much glutamate has escaped — this calcium rise drives D-serine release, widening the postsynaptic NMDA detection window. The second arm simultaneously activates presynaptic mGluR2/3 receptors via Gi, suppressing adenylyl cyclase and reducing vesicle release probability — a direct autoinhibitory brake on the very source of the overflow. These two arms run in opposite directions from the same trigger: the astrocyte brakes the presynapse while amplifying the postsynaptic learning window at the same time. The functional logic of this arrangement is also worth noting. At the moment of spillover, the synapse is already releasing at high volume — the presynapse does not need to be told to release more, it needs to be prevented from exhausting itself. Simultaneously, the postsynapse is receiving a large wavefront and is exactly the right moment to maximize its coincidence detection sensitivity. The push-pull architecture serves both needs with a single signal, without requiring any separate coordination mechanism.
The postsynapse responds to the glutamate wavefront through its AMPA receptors, depolarizing the membrane. But full calcium entry through NMDA receptors only occurs if two conditions are met simultaneously: the membrane must be sufficiently depolarized to eject the magnesium block, and D-serine released by the astrocyte must be present as a co-agonist. This is the coincidence detection step — both conditions are required, and the astrocyte's D-serine supply is what makes it a three-party coincidence rather than a two-party one.
Once the spike is complete, the astrocyte vacuums up residual glutamate via its transporter proteins, and the harder it works at this clearance the faster it runs its glycolysis engine, converting blood glucose into lactate and pumping it into the extracellular space. Both the presynapse and postsynapse absorb this lactate to power their reset pumps — refilling the vesicle pool and restoring the resting membrane potential respectively. The energy supply is therefore coupled to activity: busier synapses generate more demand on the astrocyte, which in turn fuels faster recovery.
## Intermediate scale: temporary tuning between spikes
If firing is sustained rather than isolated, the system begins temporary adjustments that do not yet commit to structural change. In the presynapse, sustained high-frequency firing keeps residual calcium elevated, progressively increasing release probability and mobilizing vesicles from deep reserve storage into the readily-releasable pool — a priming of the launchpad without permanently expanding it. If frequency is low and sparse, the reverse happens: the pool depletes faster than it refills and release probability falls.
In the postsynapse, sustained high-frequency input keeps the membrane depolarized long enough to hold the magnesium block off continuously, allowing calcium to accumulate gradually rather than in isolated pulses. This accumulation is the early signal for plasticity, but it is not yet sufficient on its own to commit a structural change.
In the astrocyte, sustained activity keeps mGluR5 activated and D-serine release elevated above its baseline pulse level, maintaining a wider NMDA detection window for as long as the high-frequency drive continues.
The neuromodulatory broadcast then sets the critical context gate. If dopamine or norepinephrine levels cross their respective thresholds — signaling that the current activity pattern is behaviorally significant — PKA activity rises and phosphorylates three targets in sequence. It lowers the threshold for AMPA receptor insertion by priming the GluA1 subunit at Ser845. It silences the LTD phosphatase PP1 by phosphorylating DARPP-32, effectively blocking the forgetting machinery from running while the save signal is present. And it translocates to the nucleus to phosphorylate CREB, enabling the gene expression needed to build new structural proteins. Acetylcholine from the basal forebrain acts in parallel, lowering the global LTP threshold — making the entire system more sensitive to incoming patterns during periods of high attention.
## Slow scale: the commit decision and structural rewriting
Once the fast and intermediate dynamics have run, the system evaluates a three-layer filter to decide whether to permanently rewrite its architecture. The first layer asks whether a genuine event occurred — specifically whether postsynaptic calcium rose above the high threshold at sufficient speed to activate the LTP kinase pathway. The second layer asks whether that event was excessive enough to saturate the cleft and trigger mGluR5 on the astrocyte. The third layer asks whether the neuromodulatory context validated the event as worth saving — whether PKA has primed the insertion machinery and silenced the forgetting machinery.
Only when all three conditions align does structural rewriting proceed, and when it does, all three compartments are rewritten simultaneously and in the same direction.
In the potentiation branch, the postsynapse anchors new AMPA receptors into the membrane via CaMKII — a process made easier because PKA has already lowered the insertion threshold — and the spine itself physically enlarges. The presynapse expands its active zone, clusters more calcium channels directly beneath the docking area to tighten the coupling between electrical events and vesicle release, and increases its baseline release probability. The astrocyte retracts its process walls inward toward the synapse, secretes structural matrix proteins to seal and stabilize the channel, upregulates its baseline D-serine synthesis for chronic NMDA priming, and reduces its glutamate clearance rate — meaning future signals will linger longer in the cleft rather than being vacuumed away immediately.
In the depression branch, the phosphatase PP1 wins instead of CaMKII, and all three changes run in reverse simultaneously. The postsynapse internalizes receptors and the spine shrinks. The presynapse removes docking slots, scatters its calcium channels away from the active zone, and pulls vesicles back into deep reserve storage. The astrocyte secretes matrix metalloproteinases to dissolve the structural scaffold, cuts its D-serine supply to starve the NMDA gate chronically, and extends its process walls outward away from the synapse, loosening the diffusion barrier so that future glutamate bleeds away faster rather than concentrating at the cleft.
If the calcium event occurred but the neuromodulatory save signal did not arrive, only transient changes happen — early receptor insertion and brief facilitation — both of which reverse within minutes without trace. If no threshold was crossed at all, nothing changes and the current structural state is simply held.
## The critical asymmetry
The astrocyte's perisynaptic wall distance is the variable that makes both outcomes self-reinforcing rather than merely additive. When it moves inward during potentiation, it concentrates glutamate at the cleft, maintains D-serine near the postsynapse, and tightens the presynaptic feedback loop — making future high-frequency events even more likely to cross the threshold. When it moves outward during depression, it dilutes the signal, starves the NMDA gate, and loosens the presynaptic feedback — making future events even less likely to reach threshold. The astrocyte therefore does not simply mirror what the neurons decide: it actively deepens the valley the synapse has already rolled into, in whichever direction that happens to be.
Dopamine is produced primarily by neurons in the Substantia Nigra pars compacta (projecting to the striatum, relevant for motor learning and habit formation) and the Ventral Tegmental Area (VTA) (projecting to the prefrontal cortex and limbic system via the mesolimbic and mesocortical pathways, relevant for reward, motivation, and the "save button" function in your model).
Norepinephrine is produced almost exclusively by the Locus Coeruleus, a tiny nucleus in the brainstem pons. Despite its small size it projects diffusely across virtually the entire brain — cortex, hippocampus, cerebellum, spinal cord. It's essentially the brain's arousal and signal-to-noise broadcaster, firing tonically at low rates during calm wakefulness and phasically during novel or stressful events.
Acetylcholine has two main sources: the basal forebrain nuclei (including the nucleus basalis of Meynert) projecting to the cortex and hippocampus — relevant for attention and learning gating — and the medial septum projecting specifically to the hippocampus, where it strongly modulates theta rhythms and memory encoding.
What's striking in the context of your model is that all three systems share the same architectural logic: a tiny, localized cell population broadcasts a global contextual signal that shifts the operational threshold of millions of synapses simultaneously — none of them carrying specific content, all of them modulating how content gets written.
Great question, and the answer is that the three scales are not strictly sequential in clock time — they overlap and nest inside each other, and their timing depends on the specific process involved.
**The intermediate scale (seconds to minutes)** happens largely *during* and *immediately after* a burst of activity, not between spike trains. Short-term facilitation and depression are running continuously within a spike train — each spike modifies the probability of the next one within the same train. The mGluR overflow sensing and D-serine escalation happen within seconds of sustained firing. The PKA/cAMP cascade triggered by dopamine or norepinephrine also runs within minutes of the neuromodulatory broadcast. So the intermediate scale is essentially the tail of the fast scale — it starts during the event and decays over the following minutes if no structural commitment is made.
**The slow scale is more complex and splits into at least two sub-phases:**
Early LTP consolidation — the initial receptor insertion and active zone expansion — begins within minutes to an hour of the triggering event, overlapping with the late intermediate scale. This phase requires protein synthesis from existing mRNAs already present at the synapse, so it can happen locally at the dendritic spine without waiting for the nucleus.
Late LTP consolidation — the full structural rewriting involving new gene expression via CREB, synthesis of new scaffold proteins, ECM secretion by the astrocyte, and permanent architectural changes — requires hours and depends on signals reaching the nucleus and new proteins being shipped back out. This is where sleep becomes genuinely important.
**Sleep plays a specific and active role** at the slow scale for two distinct reasons. First, during slow-wave sleep, the hippocampus replays compressed versions of the day's activity patterns — spike sequences that occurred during waking are reactivated at higher speed during sleep, which is thought to re-trigger the intermediate-scale cascades in a quieter, lower-noise environment where the signal-to-noise ratio is high and competing inputs are absent. This replay is believed to drive the late consolidation phase that could not complete during waking because the system was too busy processing new inputs.
Second, slow-wave sleep is also when the opposite process — synaptic downscaling — runs most aggressively. The synaptic homeostasis hypothesis proposes that waking potentiates synapses broadly and indiscriminately due to the high volume of incoming experience, and that sleep is when the LTD branch of the commit function runs systematically across the cortex, pruning weak or unvalidated potentiations back toward baseline. The astrocyte's MMP secretion and matrix dissolution likely peak during this period. This is the biological basis for the observation that sleep deprivation impairs both memory consolidation and the ability to form new memories the following day — the slow-scale commit function simply did not run.
**Active forgetting at the slow scale** has a slightly different timing. The LTD branch requires not an acute event but a chronic condition — persistent low-frequency or mistimed activity over hours to days. A synapse that simply stops being used does not immediately dissolve; it requires sustained disuse combined with the astrocyte detecting the absence of meaningful signal over an extended period before MMP secretion begins and the structural dismantling proceeds. This is why skills and memories are not lost overnight but degrade over weeks to months of disuse.
So to summarize the timing picture: the intermediate scale runs during and immediately after activity. Early structural changes begin within the hour following a validated event. The full slow-scale commit — both potentiation and pruning — runs primarily during sleep, using replay to complete the potentiation side and homeostatic downscaling to clean up the depression side. The organism's rest period is therefore not passive recovery but the primary execution window for the slow-scale branch of the commit function.Great question, and the answer is that the three scales are not strictly sequential in clock time — they overlap and nest inside each other, and their timing depends on the specific process involved.
**The intermediate scale (seconds to minutes)** happens largely *during* and *immediately after* a burst of activity, not between spike trains. Short-term facilitation and depression are running continuously within a spike train — each spike modifies the probability of the next one within the same train. The mGluR overflow sensing and D-serine escalation happen within seconds of sustained firing. The PKA/cAMP cascade triggered by dopamine or norepinephrine also runs within minutes of the neuromodulatory broadcast. So the intermediate scale is essentially the tail of the fast scale — it starts during the event and decays over the following minutes if no structural commitment is made.
**The slow scale is more complex and splits into at least two sub-phases:**
Early LTP consolidation — the initial receptor insertion and active zone expansion — begins within minutes to an hour of the triggering event, overlapping with the late intermediate scale. This phase requires protein synthesis from existing mRNAs already present at the synapse, so it can happen locally at the dendritic spine without waiting for the nucleus.
Late LTP consolidation — the full structural rewriting involving new gene expression via CREB, synthesis of new scaffold proteins, ECM secretion by the astrocyte, and permanent architectural changes — requires hours and depends on signals reaching the nucleus and new proteins being shipped back out. This is where sleep becomes genuinely important.
**Sleep plays a specific and active role** at the slow scale for two distinct reasons. First, during slow-wave sleep, the hippocampus replays compressed versions of the day's activity patterns — spike sequences that occurred during waking are reactivated at higher speed during sleep, which is thought to re-trigger the intermediate-scale cascades in a quieter, lower-noise environment where the signal-to-noise ratio is high and competing inputs are absent. This replay is believed to drive the late consolidation phase that could not complete during waking because the system was too busy processing new inputs.
Second, slow-wave sleep is also when the opposite process — synaptic downscaling — runs most aggressively. The synaptic homeostasis hypothesis proposes that waking potentiates synapses broadly and indiscriminately due to the high volume of incoming experience, and that sleep is when the LTD branch of the commit function runs systematically across the cortex, pruning weak or unvalidated potentiations back toward baseline. The astrocyte's MMP secretion and matrix dissolution likely peak during this period. This is the biological basis for the observation that sleep deprivation impairs both memory consolidation and the ability to form new memories the following day — the slow-scale commit function simply did not run.
**Active forgetting at the slow scale** has a slightly different timing. The LTD branch requires not an acute event but a chronic condition — persistent low-frequency or mistimed activity over hours to days. A synapse that simply stops being used does not immediately dissolve; it requires sustained disuse combined with the astrocyte detecting the absence of meaningful signal over an extended period before MMP secretion begins and the structural dismantling proceeds. This is why skills and memories are not lost overnight but degrade over weeks to months of disuse.
So to summarize the timing picture: the intermediate scale runs during and immediately after activity. Early structural changes begin within the hour following a validated event. The full slow-scale commit — both potentiation and pruning — runs primarily during sleep, using replay to complete the potentiation side and homeostatic downscaling to clean up the depression side. The organism's rest period is therefore not passive recovery but the primary execution window for the slow-scale branch of the commit function.
