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.

Heavy vision recurrence

90%

ResNet-152 heavy GPU calls avoided. 9.96x speedup, 100% agreement at 10x recurrence.

Llama 8B exact recurrence

90%

LLM calls avoided. 10.03x speedup, 100% agreement at 10x recurrence.

Same document, different questions

87.1%

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

Llama 8B energy

79.9%

Total board energy reduction on a 5x recurrence workload.

Beyond prefix caching

~1/5

About one fifth the energy of a prefix-cache-only path.

Meaning-level reuse

40%

Llama 8B calls avoided at a conservative threshold, with 100% reuse precision.

Taught-fact composition

100%

Invented facts composed correctly. Corrections changed future answers immediately.

Agent loop reuse

0

LLM calls on an unchanged rerun. A mid-pipeline change recomputed only affected steps.

Storage and memory overhead

16 MB

Reuse runtime beside a 6.2 GB model. 40 KB to 3 MB per 1,000 reused answers.

Embodied and scale

The same reuse core powers NeuraFrame Embodied, and it is built so that learning over time does not slow it down.

Embodied: train, ship, learn

In the Embodied test a simulated agent learned a six-situation scenario to 100% memory recall by the second epoch, then a fresh agent booted warm from the shipped memory pack and acted correctly with zero re-teaching.

Warm-start transfer

The learned memory moves as one portable file. A robot boots knowing what it was taught in simulation, with no retraining and no model on board.

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.

What the benchmarks do not claim