AI Architecture Updates: September 2, 2026

1. Context Engineering as an Architectural Decision, Not a Prompt Trick

InfoQ. Ricardo Ferreira of Redis argues that production LLM systems need a deliberate memory architecture: a TTL-bounded short-term layer for the current session and a vector-searched long-term layer, rather than one growing transcript. His central cost finding is that unbounded context growth makes API spend scale superlinearly, so he caps context by token budget instead of message count and summarizes older turns to preserve signal. The talk also covers reranking retrieved memories by relevance, semantic caching for equivalent queries, and query compression to resolve pronouns across turns, with the warning that switching models invalidates reranker score calibration. Source

2. Running Real Inference Workloads Directly in the Browser

InfoQ. The presentation covers moving real AI workloads into the browser rather than a server, and what changes architecturally when inference runs on the user’s device: no per-request inference cost, no data egress, and a hard ceiling set by whatever hardware the visitor happens to have. The privacy argument is the strongest driver, since data that never leaves the page removes an entire class of compliance work. It pairs directly with the WebGPU kernel work shipping this week, which supplies the missing performance layer. Source