Memento: Communicating Sensor Observations Through Shared Memory

1 Introduction

• Wireless communication is increasingly a bottleneck as sensing systems collect vast amounts of data over low-rate links.The measurements span scalar readings and high-dimensional modalities, including images, video, LiDAR, radar, and audio.Source: §1 · p. 1
• Unlike fidelity-oriented compression and task-oriented communication, which may be tied to fixed tasks, models, or training distributions, Memento uses retrieval over persistent memory.Memento revisits reference-based reuse for embedded sensing rather than requiring a fixed task-specific communication model.Source: §1 · pp. 1–2
• Observations often resemble content already seen by the same sensor, and non-volatile memory can retain examples as environments change.Caches and video codecs also reuse previously transmitted information, while embedded sensing faces severe bandwidth and energy constraints.Source: §1 · pp. 1–2
• Memento sends a compact reference for a similar observation already in shared memory, while the first observation is transmitted in full.The receiver uses the reference to retrieve the corresponding entry for downstream tasks.Source: Figure 1 · p. 1, §1 · p. 2
• Across four datasets, Memento traces the accuracy-payload-size and accuracy-energy Pareto frontier against baselines.The evaluation covers image and radar workloads on an MCU, SD-card, and LoRa platform.Source: §1 · p. 2
• Memento combines multi-level vector search and performance modeling with lightweight online adaptation that grows its table without encoder retraining.Under distribution shift, model-based baselines lose all accuracy while Memento retains classification performance.Source: §1 · p. 2

2 Memento

• Memento maps observations into a shared feature space, sends an index for a close match, and sends a fallback payload and inserts the sample when no close match exists.The design uses a shared embedding space in which geometric proximity corresponds to feature similarity.Source: §2 · p. 2, Figure 3 · p. 3

2.1 Vector Featurization

• A frozen pre-trained encoder produces unit-norm embeddings, optionally reduced by a random projector, with cosine similarity used to compare vectors.Both endpoints use the same feature space.Source: §2.1 · pp. 2–3

2.2 Table Construction

• Synchronized transmitter and receiver tables share an index space, so an index can retrieve the corresponding stored payload and optional label without sending the raw observation.Deployment can begin with near-empty tables or tables prepopulated from a prior over the data stream.Source: §2.2 · p. 3

2.3 Optimized Vector Table

• IVF places centroids in fast Level-1 memory and vectors in slower Level-2 memory, routing each query to a centroid-associated partition for a fine-grained search.The two-pass search first checks centroids, then searches the selected partition.Source: §2.3 · pp. 3–4
• Approximately balanced IVF partitions reduce Level-2 vectors examined from M to about M/K; a partition exceeding its capacity is split and its vectors redistributed.The split replaces one centroid with two new centroids and corresponding partitions.Source: §2.3 · p. 4

2.4 Retrieval and Transmission

• A confidence threshold determines whether the sender transmits the matched index or a fallback payload; misses append the embedding and payload to the tables.The threshold is tuned on an unlabeled validation stream to target a desired hit rate.Source: §2.4 · p. 4
• For classification, Memento evaluates top-1 accuracy using the effective prediction formed from local retrieval and cloud predictions.The communication cost mixes index-only and fallback transmissions, and expected on-wire latency depends on the link rate.Source: §2.4 · p. 4

3 When Does Memento Pay Off?

• Memento’s cost is modeled as local work plus transmission, with the hit rate p set at a threshold that meets the application’s quality requirement.Hardware-dependent inputs distinguish the costs associated with the platform and link.Source: §3 · p. 4

3.1 Query Cost Model

• Expected latency combines local computation with radio transmission, while expected energy includes encoding, lookup, communication, fallback processing, and table updates.The model accounts for fallback preparation and updates on misses, and assumes constant-rate delivery for transmission time and energy.Source: §3.1 · pp. 4–5

3.2 Break-Even Conditions

• For positive denominators, retrieval reduces latency below R⋆ T and energy below R⋆ E, where the thresholds balance transmission savings against added local costs.The model compares retrieval with direct transmission and includes avoided fallback preparation on hits.Source: §3.2 · p. 5

3.3 Using the Model to Analyze Tradeoffs

• The modeled link cases use measured Apollo4 Plus and LoRa parameters plus documented rates and power for NB-IoT, BLE, and Wi-Fi; setup, receive windows, and retries are omitted.Radio cases are illustrative continuous-transfer approximations.Source: §3.3 · p. 5, Table 1 · p. 5
• At 25% hits, Memento reduces latency by approximately 25% for LoRa and 20% for NB-IoT, and reduces energy by approximately 25% for both.Retrieval can pay off at modest hit rates on low-throughput links.Source: §3.3 · p. 5
• At 90% hits, Memento reduces latency by approximately 90% for LoRa and 85% for NB-IoT, and reduces energy by approximately 90% for both.The figure compares predicted latency and energy costs with direct transmission as hit rate varies.Source: §3.3 · p. 6, Figure 4 · p. 6
• BLE reduces latency only above an 81.6% hit rate and does not save energy, while Wi-Fi saves energy above 49.2% but increases latency even at a perfect hit rate.Higher throughput leaves less transmission time to save; energy savings also depend on sender transmission power.Source: §3.3 · p. 6

4 Implementation

• The image and radar encoders are trained through knowledge distillation, with DINOv2 and ConvNeXt-Tiny teachers and MCUNet and RadarTCN students, respectively.The image student produces 128-dimensional embeddings; the radar student produces 64-dimensional embeddings.Source: §4 · p. 6
• The prototype uses an Ambiq Apollo4 Plus evaluation board at 192 MHz, TensorFlow Lite Micro for neural-network execution, and removable microSD for the vector database.The database stores 50,000 128-dimensional vectors with 512 coarse clusters; the neural network uses a 600 kB SRAM tensor buffer.Source: §4 · p. 6
• At startup, IVF routing metadata is loaded into SRAM while database vectors and labels remain on the SD card; search reads only the selected partition.The selected partition is read from a contiguous file, allowing online updates per file without affecting other partitions.Source: §4 · p. 6
• Retrieval averages 3.4 ms for centroid lookup and distance computation in SRAM, 38.1 ms to read the selected microSD partition, and 1.1 ms for within-partition distances.SD card access accounts for most measured retrieval latency across 10 measurements.Source: §4 · p. 6
• Wireless transmission uses a Semtech LR2021 LoRa radio at 868.1 MHz, with a 125 kHz bandwidth, spreading factor 9, and coding rate 4/5.Payloads are capped at 255 B, with larger messages fragmented.Source: §4 · p. 6

5.1 Experimental Setup

• The evaluation covers three image workloads and one radar workload, testing OPT and QUAL variants on task accuracy, transmitted bits, MCU latency, and energy.The reported image metric is top-1 accuracy; CARRADA uses binary accuracy measured as mean agreement over four presence bits.Source: §5 · p. 6, §5.1 · p. 7

5.2 Accuracy vs. Bandwidth Savings

• Memento contributes accuracy-payload Pareto-frontier points on every dataset, including operating points with one to three orders of magnitude less communication than cloud classification.Examples include OPT at 96.04% accuracy with a 30.7× payload reduction on CCT-38 and QUAL at 96.22% with a 16.6× reduction on UA-DETRAC.Source: Figure 5 · p. 7, §5.2 · p. 8

5.4 Energy Analysis

• At high accuracy, QUAL matches CloudCLS on CCT-38, UA-DETRAC, and CARRADA with 5.5 to 27.1× less energy, and comes within 0.14 points on CCT-All with 14.7× less energy.OPT uses 1.5 to 2.4× less energy than LocalCLS, with an accuracy cost of 0.25 to 7.92 points.Source: Figure 6 · p. 8, §5.4 · p. 9

5.3 Latency Analysis

• Memento’s OPT requires 299 ms of local computation and 1.27 s of mean airtime, achieving the lowest compute-plus-airtime cost at 1.57 s.It sends a six-byte reference for 91.9% of queries and uses JPEG fallback for the remaining 8.1%.Source: Figure 7 · p. 8, §5.3 · p. 9

5.6 Qualitative Analysis

• On two CCT-All queries, a six-byte Memento reference lets the receiver inspect a stored image with visual details that a class label or compressed payload may lose.The query and reference images are 21 and 79 days apart, respectively.Source: Figure 9 · p. 10, §5.6 · p. 10

5.5 Online Adaptation

• After a class shift, Memento stabilizes at 83–87% accuracy on unseen classes while transmitting 24–38 KiB per query; its encoder remains fixed and only the vector table grows.The local classifier, LimitNet, and AgileNN drop to 0% top-1 accuracy, while cloud classification and Progressive JPEG maintain accuracy at high communication cost.Source: Figure 8 · p. 9, §5.5 · p. 10

6 Related Work

• The related work spans image compression, region-of-interest detection, edge offloading, and data-driven wireless communication, including DeepJSCC and compressive offloading methods.The paper reports that the cited methods in,, and are outperformed by the evaluated baselines [12] [14].Source: §6 · p. 11
• AgileNN and LimitNet are identified as the most relevant comparisons; Memento is reported to outperform them in latency, energy efficiency, and communication overhead across several datasets.LimitNet uses content-aware progressive encoding, while AgileNN offloads neural-network inference.Source: §6 · p. 11
• Related approaches are described as often relying on fixed datasets, computationally expensive models, or limited adaptability to deployment scenarios.The paper presents Memento as a lightweight, adaptive framework that exploits redundancy in real-world sensor data.Source: §6 · p. 11

7 Discussion

• Memento performs best in deployments with temporally redundant observations; when observations rarely repeat, full-sample fallback becomes more frequent and communication savings decline.The fallback path maintains correctness in those dynamic environments.Source: §7 · p. 11
• The vector table grows with insertions, but the paper reports retrieval performance saturating at relatively modest table sizes and suggests bounding growth with eviction, clustering, or redundancy pruning.The discussion notes that modern IoT devices commonly include gigabytes of non-volatile storage.Source: §7 · p. 11
• Memento assumes synchronized sender and receiver tables, and retrieval effectiveness depends on embedding quality; the discussion identifies image and radar as evaluated modalities and other modalities as potential applications.Suggested synchronization mechanisms include acknowledgments for fallback transmissions or periodic consistency checks.Source: §7 · p. 11

• Memento uses vector retrieval over persistent memory to send compact references for many observations instead of transmitting them in full.Source: §8 · p. 11
• An efficient multi-level retrieval structure keeps large vector tables in persistent memory with negligible lookup overhead on resource-constrained hardware.Source: §8 · p. 11
• Across three image datasets and one radar dataset, Memento outperforms representative compression-based and task-oriented baselines on accuracy-bandwidth and accuracy-energy tradeoffs.Source: §8 · p. 12
• The results suggest shifting work from communication to local computation as sensing workloads grow in scale and complexity.Source: §8 · p. 12