Carousel Memory Across Domains

• Carousel memory is a concept denoting the retention and reuse of structured history in diverse fields, from serialized user journeys in recommender systems to spark-pattern persistence in pulsar studies.
• In recommender systems and click models, carousel memory involves serializing user interactions and exposure histories to improve prediction accuracy and ranking performance, as evidenced by enhanced HR and NDCG metrics.
• In continual learning and pulsar astrophysics, carousel memory refers respectively to a hierarchical replay mechanism for mitigating forgetting and to the physical persistence of rotating spark patterns across emission heights.

Carousel memory is a domain-dependent term used to denote the retention and operational use of structured history in systems organized around carousel-like entities. In recommender systems, it refers to a serialized memory of cross-surface user interactions with carousels, items, and search events, operationalized as next-token prediction in a shared sequence model (Salle et al., 22 May 2026). In click modeling, it denotes dependence on ordered panel, page, slot, topic, exposure, and click history in carousel interfaces (Kang et al., 23 Jun 2025). In continual learning, it names a hierarchical episodic-memory design that rotates samples between RAM and storage to mitigate catastrophic forgetting (Lee et al., 2021). In pulsar astrophysics, it refers to the persistence of a rotating spark-carousel pattern across emission heights and observing frequencies, recoverable through geometry-induced phase offsets in subpulse modulation (Maan, 2018). These usages are technically unrelated, but each centers on retained structure that reappears in later inference, prediction, or measurement.

1. Terminological scope and common abstraction

The expression “carousel memory” does not designate a single canonical construct across the literature. Instead, it appears in multiple research areas with different objects of study, different mathematical formalisms, and different operational goals. In streaming discovery, the relevant object is the user’s cross-surface interaction history, serialized into a “user story” sequence containing attributes, sessions, watch events with surface and carousel context, and search events (Salle et al., 22 May 2026). In interface modeling, the object is the ordered interaction history over vertically stacked horizontal lists, formalized through dependencies on topics $T$, displayed items $Y$, and prior clicks $C'$ (Kang et al., 23 Jun 2025). In continual learning, the object is a replay buffer hierarchy spanning RAM and persistent storage, with asynchronous sample rotation (Lee et al., 2021). In pulsar studies, the object is the rotating spark pattern on the polar cap, whose imprint is retained by plasma subbeams and re-sampled at different frequencies and heights (Maan, 2018).

A plausible unifying interpretation is that carousel memory always denotes persistence under structured re-exposure. What is retained differs by field: behavioral context in recommendation, exposure and click history in interface models, training examples in continual learning, and spark-pattern geometry in pulsar emission. The term “memory” is therefore not metaphorically uniform: in some cases it is an explicit data structure, in others a probabilistic conditioning set, a systems mechanism, or a physical imprint.

2. Cross-surface behavioral memory in unified ranking

In TubiFM, carousel memory is the unified, tokenized account of how a viewer encounters, engages with, and responds to carousels across discovery surfaces (Salle et al., 22 May 2026). Rather than maintaining separate logs or features for item ranking, carousel ranking, and search, the system serializes the entire cross-surface journey into a single “user story” sequence. The fixed linearization is attributes $\rightarrow$ sessions $\rightarrow$ events in chronological order. A watch event records the surface, the carousel identity, the watched item identifier, viewing duration, and temporal context including day-of-week, hour, session markers, and elapsed inter-session time. Search events are recorded for every query, including search-as-you-type intermediates and queries that do not lead to a watch. The sequence is simultaneously the retained memory and the interface by which downstream ranking tasks are posed.

The token schema interleaves the pretrained Llama 3.2 1B tokenizer’s BPE vocabulary with new domain-specific atomic tokens. The grammar includes structural tokens such as <|begin_sessions|> and <|session|>, event tokens <|watch|> and <|search|>, surface tokens such as <|surface=home|> and <|surface=search|>, carousel tokens <|carousel(name)|> and <|carousel(MASK)|>, atomic item identifier tokens of the form <|id(SYN###|Title)|>, and reserved unknown item and carousel tokens. An auxiliary catalog corpus ties domain tokens back to text, with user-story and catalog corpora sampled at a $20\!:\!1$ ratio.

All ranking tasks are trained as next-token prediction under the standard autoregressive cross-entropy objective

$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$

Task prompts alter only the head tokens that define what should be predicted next. At inference, a single forward pass scores candidates by the logit of their token at the next-token position. For carousel ranking,

$$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$

and for item ranking,

$$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$

The same story can therefore be reused for carousel ordering, item ranking, or search ranking simply by changing the prompt suffix.

The memory contribution of the carousel-specific fields is explicit. Watches are tagged with their source carousel and surface, enabling the model to associate carousel identities with engagement signals such as viewing duration. Queries that do not lead to watches remain in memory as negative search exposure signals. The schema does not serialize pure impression logs for non-clicked items within a carousel; exposure for carousels is therefore represented primarily through watched associations, while search retains negative signals through non-converting queries. To mitigate container bias, TubiFM replaces the observed surface+carousel pair with <|surface=home|><|carousel(MASK)|> with probability $0.1$ during training, and uses an item-UNK rate of $Y$0 for robustness to new catalog items.

The reported empirical evidence links this serialized carousel memory to both ranking quality and systems simplification. TubiFM uses Llama 3.2 1B with context length 1024 tokens and average serialized story length approximately 560 tokens; training data comprise approximately 20M viewers, approximately 800M watches, approximately 66M searches, approximately 100K titles, and approximately 1K carousels across four primary surfaces. In offline evaluation, unified TubiFM outperforms specialist baselines across item, carousel, and search ranking. For carousel ranking against the strongest baseline, HSTU, HR@8 is $Y$1 versus $Y$2 and NDCG@8 is $Y$3 versus $Y$4; HR@50 and HR@100 are $Y$5 and $Y$6. Online A/B tests report search TVT $Y$7, carousel TVT $Y$8, and item TVT $Y$9 with item ranking statistically neutral. Across all three tasks, TubiFM serves on L40S GPUs and reduces p99 ranking latency from approximately $C'$0 to approximately $C'$1 (Salle et al., 22 May 2026).

This formulation places carousel memory at the center of a unified discovery stack. The significance lies less in a new standalone memory module than in the use of one serialized behavioral history as a common conditioning substrate for multiple tasks that are usually siloed.

3. Carousel interfaces and click-model memory

In click modeling, carousel memory is defined through the dependence structure of click probability in interfaces composed of multiple titled horizontal lists stacked vertically (Kang et al., 23 Jun 2025). Unlike single-list search, where examination is typically modeled as top-to-bottom sequential with position bias, carousel interfaces require joint modeling of vertical panel selection and horizontal paging or swiping. This induces panel/page bias, topic-conditioned interactions, sequential dependence, and revisitation. The memory-bearing variables include carousel topics $C'$2, items $C'$3 at panel $C'$4, page $C'$5, and slot $C'$6, prior clicks $C'$7, dwell times, scroll events, revisit indicators, and session metadata.

The paper organizes click-model design around three key choices: global dependencies, sequentiality, and factorization. Global dependencies specify which observed variables influence click probability, choosing a subset $C'$8 and modeling

$C'$9

Memory is explicitly enabled when $\rightarrow$0 includes $\rightarrow$1; topic-conditioned memory requires $\rightarrow$2; exposure memory within a carousel requires $\rightarrow$3. The taxonomy includes Random $\rightarrow$4, Topics-Only $\rightarrow$5, Items-Only $\rightarrow$6, Items-Clicks $\rightarrow$7, Topics-Clicks $\rightarrow$8, Topics-Items $\rightarrow$9, and Fully Dependent $\rightarrow$0.

Sequentiality determines which subsequences condition each click. Within a carousel, this may take the form $\rightarrow$1; across carousels, topic-order memory may be encoded as $\rightarrow$2. Factorization determines how these variables combine. The paper contrasts PBM-like multiplicative decompositions with memory-aware forms such as a Topic $\rightarrow$3 Item $\rightarrow$4 Click-history product $\rightarrow$5, or attention-based mixtures over history.

For carousel click modeling with horizontal pages, the session history is formalized as $\rightarrow$6, the sigma-field of past observables up to location $\rightarrow$7. This includes topics $\rightarrow$8, exposed items, prior clicks with earlier timestamps, dwell time $\rightarrow$9, scroll events $20\!:\!1$0, revisit counts, and session metadata. A general memory-aware factorization decomposes clicks into examination, relevance, and click decision:

$20\!:\!1$1

A compact history-conditioned click probability is

$20\!:\!1$2

The examination component can encode panel, page, and slot bias together with memory terms:

$20\!:\!1$3

where $20\!:\!1$4 is panel/topic bias, $20\!:\!1$5 page bias, and $20\!:\!1$6 slot bias. The history features $20\!:\!1$7 may include count_seen$20\!:\!1$8, time_since_seen$20\!:\!1$9, sum_dwell$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$0, last_click_pos$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$1, clicked_in_topic$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$2, and vertical and horizontal scroll events. This formalism makes “carousel memory” a probabilistic dependence on ordered interface history rather than a stored representation in the systems sense.

The paper also derives a novel Topics–Items design that includes exposure memory and topic-order memory but excludes click-history memory:

$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$3

with sequentiality

$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$4

Two factorizations are given:

$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$5

and

$$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$6

This distinguishes exposure-memory models from Fully Dependent models such as CCM1, which condition on $$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$7, $$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$8, and $$L = -\sum_{t=1}^{T} \log p(x_t \mid x_{<t}; \theta).$$9.

Neural versions replace explicit PGM factorization with hidden-state or attention-based sequence modeling. An RNN formulation orders navigation events by timestamp,

$$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$0

while a transformer uses attention over prior history,

$$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$1

In these models, memory is carried by hidden states or attention weights over prior exposures, clicks, dwell, and revisitation (Kang et al., 23 Jun 2025).

A common misconception is that carousel interfaces can be treated as minor variants of single-list ranking. The taxonomy argues against that view by making panel/page bias, topic-order effects, and revisitation first-class components of the conditioning structure. Another misconception is that click memory must always be click-history memory; the Topics–Items design shows that exposure memory and topic-order memory can be modeled without conditioning on $$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$2.

4. Hierarchical episodic memory in continual learning

In continual learning, Carousel Memory, abbreviated CarM, is a hierarchical episodic-memory management strategy designed for rehearsal-based continual learning under tight RAM budgets (Lee et al., 2021). The problem setting is a non-i.i.d. stream of tasks $$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$3, where standard rehearsal-based methods maintain a small episodic memory in RAM and replay stored samples during training. Conventional methods discard samples that overflow the RAM buffer permanently. CarM instead exploits the device memory hierarchy—small, fast RAM and larger, slower persistent storage—to preserve past experiences in storage and rotate them back into RAM asynchronously.

The architecture has two principal buffers: EM, a RAM buffer supplying replay samples, and ES, a storage buffer holding all or most past samples. It also includes a stream buffer for incoming data, fetch workers that pre-process training bundles, a swap worker that decides which in-RAM samples to replace and issues storage I/O, a gate function and policies that control swapping, and optional EM partitioning to parallelize reads and writes. The execution pipeline contains five stages: data incoming, training, EM updating, storage updating, and storage sample retrieving.

The training objective is written as

$$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$4

where $$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$5 during training, and CarM continually refreshes EM from ES so that replay support expands over time. The central systems mechanism is asynchronous overlap of retrieval $$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$6 with training $$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$7, avoiding the slowdown of a synchronized sequence $$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$8.

The gate function determines which samples are swapped:

$$s(c \mid \text{story}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{carousel}(c)\mid\!\rangle),$$9

where $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$0 is a score and $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$1 controls the fraction replaced. Three policies are reported. Random marks a random $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$2 proportion of the mini-batch to be swapped. Entropy keeps hard samples and swaps easy ones using

$$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$3

where $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$4, $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$5 is prediction entropy, and $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$6 is maximum entropy. Dynamic uses Random in early passes and Entropy in later passes. When ES is full, storage eviction is random per class to maintain class balance.

CarM is presented as a drop-in memory-management layer compatible with seven rehearsal-based methods: ER, iCaRL, TinyER, BiC, GDumb, DER++, and RM. It does not alter the model architecture or core loss, except that the paper studies distillation reweighting for methods such as iCaRL. The modified iCaRL objective is

$$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$7

and the reported empirical result is that small $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$8 values, specifically $$s(i \mid \text{story}, \text{context}) = \operatorname{logit}(\text{next} = \langle\!\mid \text{id}(i)\mid\!\rangle).$$9 to $0.1$0, work best with CarM.

The empirical results characterize CarM as a memory-capacity extension with limited efficiency cost. Across evaluated datasets, final average accuracy improvements are substantial. On the CIFAR subset, DER++ improves from $0.1$1 to $0.1$2 with CarM-50 and $0.1$3 with CarM-100, while forgetting decreases from $0.1$4 to $0.1$5–$0.1$6. On Tiny-ImageNet, DER++ rises from $0.1$7 to $0.1$8 with CarM-50, the largest reported gain at $0.1$9 percentage points. On ImageNet-1000, RM improves from $Y$00 to $Y$01 and $Y$02, and DER++ from $Y$03 to $Y$04 and $Y$05. Async CarM preserves near-original training efficiency, whereas Sync serialization incurs large slowdowns; for example, DER++ shows $Y$06 Async overhead versus $Y$07 Sync overhead on CIFAR/ImageNet (Lee et al., 2021).

Here, carousel memory does not refer to a carousel user interface or a rotating physical pattern. It designates rotation of replay samples through a memory hierarchy. The naming foregrounds recurrence and reintroduction rather than spatial layout. This suggests that, in continual learning, the crucial semantic content of the term is not “carousel” as UI element but “carousel” as controlled circulation.

5. Persistent spark patterns in pulsar radio emission

In pulsar astrophysics, carousel memory refers to the persistence and recoverable mapping of a rotating spark pattern on the polar cap across different emission heights and observing frequencies (Maan, 2018). In the carousel model, regularly spaced sparks in the polar acceleration gap circulate around the magnetic axis due to the $Y$08 drift. Each spark seeds a plasma subbeam flowing along magnetic field lines, and the geometry of these subbeams retains the spark pattern as they propagate to radio-emission altitudes. Different frequencies then sample the same underlying pattern at different carousel rotation phases.

The basic quantities are the observed subpulse modulation period $Y$09, the true modulation period $Y$10, the carousel circulation period $Y$11, and the number of sparks $Y$12, related by

$Y$13

where $Y$14 is the alias order. The viewing geometry is parameterized by magnetic inclination angle $Y$15, impact parameter $Y$16, $Y$17, emission-cone radius $Y$18, magnetic azimuth $Y$19, and pulse longitude $Y$20 relative to the fiducial plane $Y$21.

The magnetic azimuth is

$Y$22

The modulation phase envelope is

$Y$23

For simultaneous observations at two frequencies $Y$24, the expected inter-frequency phase offset is

$Y$25

where $Y$26, $Y$27, and

$Y$28

For typical altitude differences of a few hundred km and $Y$29 of seconds, $Y$30 is reported as negligible.

The “memory” resides in the fact that the same spark-carousel structure is preserved across emission heights, so the observer measures correlated modulation with predictable phase offsets. This permits aliasing resolution through observed quantities alone:

$Y$31

where $Y$32 is the fractional delay between fluctuation sequences measured at two frequencies. The method does not require $Y$33, $Y$34, or $Y$35.

The paper’s case study of PSR B1237+25 uses a 327 MHz pulse sequence and analyzes two subsequences, S#11 and S#12. The observed modulation periods are $Y$36 and $Y$37. Cross-correlation of $Y$38-folds yields phase offsets of approximately $Y$39–$Y$40 for components I vs II and approximately $Y$41 to $Y$42–$Y$43 for IV vs V. Cross-spectrum analysis gives $Y$44 for S#11 and $Y$45 for S#12 on I vs II, implying phase offsets consistent with the direct measurements. The aliasing equation yields $Y$46, so $Y$47–$Y$48 is unaliased. With aberration-retardation-corrected component longitudes and geometry $Y$49, $Y$50, the paper estimates $Y$51, $Y$52–$Y$53 broadened to $Y$54–$Y$55 allowing $Y$56 in $Y$57, and hence $Y$58–$Y$59 (Maan, 2018).

The decisive observational signature is that inner and outer conal components share a common $Y$60 and show phase-locked modulation with equal magnitude and opposite sign across the fiducial plane. The paper interprets this as firm evidence that both cones originate from the same carousel of subbeams at different heights. In this usage, carousel memory is therefore a physical persistence property of the magnetospheric emission structure, encoded in measurable phase relations.

6. Comparative interpretation, misconceptions, and open distinctions

The four research usages differ in ontology, observables, and operational form. In TubiFM, carousel memory is a serialized behavioral sequence used directly for prediction across ranking tasks (Salle et al., 22 May 2026). In carousel click models, it is a dependency structure over panel, page, slot, topic, exposure, and click history (Kang et al., 23 Jun 2025). In CarM for continual learning, it is a hierarchical storage-and-replay mechanism (Lee et al., 2021). In pulsar studies, it is the persistence of a spark pattern under propagation and multi-frequency sampling (Maan, 2018). Treating these as a single concept would be misleading.

Several misconceptions are clarified by the source material. First, in recommendation settings, carousel memory does not mean complete exposure logging. TubiFM explicitly states that pure impression logs for non-clicked items within a carousel are not serialized; the strongest exposure signals are watches and searches themselves, although non-converting search queries are retained as negative signals (Salle et al., 22 May 2026). Second, in click modeling, memory is not synonymous with prior clicks. The taxonomy makes clear that memory can arise from exposure history $Y$61, topic order $Y$62, or click history $Y$63, separately or jointly (Kang et al., 23 Jun 2025). Third, in continual learning, Carousel Memory is not a new replay loss; it is a memory-management layer that leaves baseline learning rules largely intact, apart from optional retuning such as lowering distillation weight (Lee et al., 2021). Fourth, in pulsar astrophysics, carousel memory is not a generic observational correlation; it is specifically tied to the preservation of a rotating spark pattern and to geometry-induced phase offsets predicted by the carousel model (Maan, 2018).

A plausible cross-domain implication is that “carousel memory” tends to emerge where later decisions depend on structured recurrence rather than isolated events. Yet the mechanism of recurrence varies sharply. Recommendation and click modeling focus on user-interface traversal and engagement history; continual learning focuses on sample circulation under hardware constraints; pulsar studies focus on the re-sampling of a persistent physical configuration. The shared phrase therefore names an abstract family resemblance—retained structured history—but not a transferable method.

Across domains, the term is most informative when read locally, within the modeling assumptions of the specific paper. In practice, that means distinguishing between tokenized event histories, conditional-dependence graphs, hierarchical replay systems, and geometric phase-memory in radio emission, even though all are described as forms of carousel memory.