AIe deploys as a gateway: point your traffic at it, add one line to your engine configuration, keep everything else. Your models, your GPUs, your prompts — unchanged.
Every reused answer is verified before it is returned: in testing, 100% of verified answers were byte-identical to the unmodified engine's output, across three model families. Where a request truly differs from anything in memory, it is computed in full. Same computation. Same output. Less work.
No similarity threshold ever decides what your users receive — the failure mode of semantic caches, where two questions that look alike but differ in a date, a clause number or a dose are silently merged. How AIe achieves this is the subject of Navyra's patent filings; what it achieves is on the Evidence page, and reproducible in your own evaluation.
| Question | Answer |
|---|---|
| Do I change my model? | No |
| Do I retrain anything? | No |
| Do I change my prompts? | No |
| Does it replace vLLM? | No |
| Does it replace prefix caching? | No — it adds to it |
| Does traffic leave my environment? | No |
| What happens on a miss? | Normal inference, ~1% overhead |
| Can I switch it off? | Yes — instantly |
Per-tenant memory with strict isolation — one memory per customer, never shared between them. Data locality by design: nothing depends on an external API, and the whole system deploys in your VPC, on your premises, or fully air-gapped. Security architecture documentation is available under NDA.
| Inference engine | vLLM — available now |
| Other engines | Adapter roadmap — the memory is engine-agnostic |
| GPUs | NVIDIA, validated on H100 |
| Open-weight dense models | Validated — three model families |
| Frontier-scale & mixture-of-experts | In development — measured, not yet validated |
Updated as validation grows — what's listed as validated has been measured and can be reproduced under NDA.
The memory survives heavy load and full restarts: entries still answer in 0.62 ms after 574,000 tokens of unrelated traffic, where the engine's own cache had long forgotten; 89.5% of the memory returns after a restart. The memory scales to millions of entries with millisecond lookup, and is built to hold months of traffic.