Measured benchmarks on real edge AI workloads

Benchmarks were measured on local edge hardware using real heavy models including ResNet-152 and Llama 3.1 8B quantized. Implementation details are intentionally omitted. These figures come from a modest edge device on purpose: the same reuse layer runs on x86_64 servers and multi-GPU data-center hosts, where each avoided model call costs far more, so the savings scale up with your hardware.

Safe where a cache is wrong

0

Stale answers served over real changes on a recurring vision stream, while sending about 1% of the pixels. A tuned whole-image cache on the same stream served 6 to 13. A cache cannot be both safe and thrifty; a memory that understands change can.

Continuous vision on video

~1%

Pixels sent to the model on a panning stream: 3 calls across 13 frames, zero stale answers over a real change, where a whole-image cache cannot follow a thing across the pan.

Same document, different questions

87.1%

Prompt-token reduction with 100% factual accuracy and 100% semantic agreement.

Meaning-level reuse

40%

Llama 8B calls avoided on same-meaning, different-words questions at a conservative threshold, with 100% reuse precision. Byte-matching caches score zero here by definition.

Taught-fact composition

100%

Invented facts composed correctly. Corrections changed future answers immediately.

Beyond prefix caching

~1/5

About one fifth the energy of a prefix-cache-only path, because whole answers are reused, not just prompt prefixes.

Llama 8B energy

79.9%

Total board energy reduction on a 5x recurrence workload.

Storage and memory overhead

16 MB

Reuse engine runtime beside a 6.2 GB model. 40 KB to 3 MB per 1,000 reused answers. The full service (API, gateway, and engine) idles around 70 MB, and the install preflight's 8 GB floor is sized for running a model beside it (16 GB for the 8B tier).

The recurrence baseline

90%

At 10x recurrence the exact layer avoided 90% of calls (9.96x on ResNet-152, 10.03x on Llama 8B, 100% agreement), and an unchanged agent rerun needed zero. This is the table-stakes result any competent cache achieves; it is listed to show the layer adds no cost, not as the differentiator. The results above are the ones a cache cannot produce.

Scale over time

A memory is only useful if it stays fast as it fills up. The engine is built so that months of learning do not become a tax on every answer.

Does not slow down as it learns

Verified exact answers and corrections are a hash lookup, constant time no matter how large the memory grows. Semantic memory is bounded, so scan time stays flat as it learns over months and years.

Warm start

The verified memory moves as one portable file. A new machine boots knowing what another was taught, with no retraining and nothing to re-derive.

Survives restarts

Durable memory lives outside the versioned install, so upgrades, reinstalls, and reboots preserve everything the device has learned.

What the benchmarks do not claim