Sensors stop re-uploading similar scenes: on a hit, six bytes retrieve the stored image
Synopsis
Memento sends sensor observations by reference: the device searches a vector table in persistent memory for a similar past sample and, on a hit, transmits only an index, and modeling shows that on low-rate links such as LoRa a 25% hit rate cuts latency and energy by roughly a quarter.
Interpretation
A wireless sensing device can reduce one observation to a short reference instead of sending the whole image or radar frame. Earlier compression and task-oriented transmission encoded each observation independently and often relied on fixed models that needed retraining when the data distribution shifted. Across three image datasets and one radar dataset, Memento's OPT reaches 96.04% accuracy on CCT-38, 0.59 percentage points below cloud classification, at a 30.7x payload reduction; at a smaller budget QUAL reaches 92.74% at a 988x reduction, 3.15 points above local classification.
Sender and receiver each keep a table under a shared index, the sender holding vectors and the receiver holding the corresponding raw samples. Retrieval uses an inverted file structure that keeps a small centroid table in fast memory and the full vector table in slower, larger storage, so a query first picks a cluster and then compares only within it. On an Ambiq Apollo4 Plus prototype with a microSD card and a LoRa radio, the table holds 50,000 vectors of dimension 128 with 512 coarse clusters; centroid lookup and distance computation in SRAM take 3.4 ms, reading the selected partition takes 38.1 ms, and in-partition distance computation takes 1.1 ms.
Perspective
The mechanism targets deployments where observations are temporally redundant, such as camera traps, environmental monitoring or industrial inspection; in highly dynamic environments where observations rarely repeat, the system falls back to full transmissions more often, reducing communication savings while the fallback path preserves correctness. The evaluation covers three image datasets and one radar dataset, with a prototype built on an Ambiq Apollo4 Plus microcontroller, a removable microSD card and a Semtech LR2021 LoRa transceiver.
The system assumes the sender and receiver keep consistent vector tables, and practical deployments need acknowledgments or periodic consistency checks to maintain synchronization; if divergence occurs, the system recovers when a query misses and the full observation is retransmitted and reinserted. Retrieval effectiveness depends on the quality of the embedding space, since stronger encoders raise the probability that similar observations map to nearby vectors. The table grows as new samples are inserted, and the evaluation shows retrieval performance saturating at relatively modest table sizes, so deployments could bound table size with age-based eviction, clustering or redundancy pruning.