Pervaziv AI adds 3-tier inference cache to Cortex

Pervaziv AI introduced a 3-tier Cortex Inference Cache Architecture on Sept. 22, 2026, aimed at speeding enterprise AI workflows without weakening access control or freshness checks. The company says the design can reuse context, prompt prefixes or exact responses, depending on what remains valid, with measured gains reaching up to 150x faster prefill and 11x faster response delivery in tested workloads.

Why it matters: - Enterprise AI often repeats work across long, multi-step tasks. Cortex is trying to cut that repetition without reusing stale context, outdated permissions or incorrect results. - The architecture is meant to separate speed gains from correctness checks, which matters for workflows that move across browser, code, cloud and live data. - Pervaziv AI is positioning caching as a governed capability, not just a performance feature.

What happened: - Pervaziv AI announced a 3-Tier Cortex Inference Cache Architecture on Sept. 22, 2026. - The architecture adds three forms of reuse: application context, model prompt prefix and exact completed response. - Cortex is designed to decide which tier is valid for the current request based on source state, permissions and execution context.

The details: - Tier one reuses eligible application context when the source material and authorization are still current. - Tier two reuses an identical model prompt prefix when the beginning of the request matches under the correct isolation boundary. - Tier three reuses a completed response only for approved, exact-match read-only routes. - A repeated conversation-style workload showed prompt processing falling from 2,913 milliseconds cold to 19.3 milliseconds warm, about 150x faster. - A separate inference test showed prompt processing falling from 2,693 milliseconds cold to 19.3 milliseconds warm, about 140x faster. - In the repeated conversation-style run, warm total latency stayed within about 6% across repeat requests. - In a separate approved read-only text route, an initial request finished in 3,132 milliseconds and an identical repeat returned a cache hit in 281 milliseconds. - That repeat was about 11x faster and about 91% lower in end-to-end latency. - A workload matrix completed 240 requests with zero request failures. - For one mid-length workload under moderate concurrency, throughput reached 2.25x. - At larger inputs, extra concurrency produced little throughput benefit and worse tail latency. - Pervaziv AI says these measurements describe different parts of the system, not one universal speed claim. - The company says prompt-prefix reuse speeds repeated input processing while the model still generates a new answer. - The company says exact response reuse can avoid generation only for an approved identical request. - Cortex treats freshness as a relationship among the request, inputs, current authorization and execution configuration. - Changes to documents, branches, browser pages, tool contracts, model behavior or policy can make earlier work ineligible. - Cortex distinguishes invalidation, which blocks selection of stale state, from eviction, which frees capacity. - If safe reuse is unavailable or uncertain, Cortex is designed to take a cold path and recompute the work. - The platform evaluates preparation time, time to first token, generation rate, end-to-end latency, throughput, tail latency, memory pressure and fairness together. - Pervaziv AI is also tying the cache architecture to existing Cortex products, including Cortex Connect, Cortex Cloud and Cortex Discover. - Cortex Connect spans mobile, browser experiences and Visual Studio Code. - Cortex Cloud provides durable managed execution for eligible work. - Cortex Discover adds an agentic AI browser with current page awareness, selected tab context, visual understanding and supported actions. - The company describes Cortex as an Enterprise AI Control Layer for secure, governed AI work across browser, mobile, development and cloud environments.

Between the lines: - The design is a response to a common enterprise AI problem: repeated work is expensive, but blind reuse can create security and correctness risks. - The architecture draws a hard line between context reuse, model-input reuse and full response reuse, which should make performance claims easier to evaluate. - The emphasis on exact matches and explicit route eligibility suggests Pervaziv AI is prioritizing control over aggressive cache hit rates. - By making cache policy route-specific, Cortex can support read-only analysis differently from coding, security investigation or workflows that affect external systems.

What's next: - Pervaziv AI said candidate deployments can be evaluated independently for correctness, privacy, cold and warm behavior, source and permission invalidation, memory use, mixed workloads, failover, fairness and tail latency. - The company is treating the three tiers as independently enableable, so one reuse path can be deployed without assuming the others are appropriate. - Broader rollout will likely depend on whether enterprise teams accept the tradeoff between speed gains and stricter reuse eligibility.

The bottom line: - Cortex is being built to reuse only what is still safe to reuse, and to recompute everything else.