Research
Writing from the VLabs engineering team. We write about how our own system actually works and about where AI as a whole is going, with real measurements, cited sources, and what does not work yet.
-
How AI systems remember: the architecture of machine memory in 2026
A survey of how memory is built in AI assistants and agents. Products converged on plain text files with background consolidation, simple baselines matched dedicated memory systems in eight published comparisons, and forgetting, compaction and memory poisoning remain unsolved.
-
How long can an AI work on its own? Reach, reliability and the limits of measurement
An analysis of public data to November 2026. The length of task AI agents can complete has doubled every four months. Across 23 models, the length they complete reliably has stayed about five times shorter for three years, and the instruments that measure both stopped working in 2026. Includes what has been measured for the current generation: GPT-6 Astra, Gemini 4 Argon, Claude Fable 5.1 and Claude Opus 5.5.
-
After you press Send: anatomy of one Vera work turn
One request followed through eight stages, from the moment the server accepts it to the moment the result is in your hands: a server-side ledger, a browser that runs on your own device, a sandbox with no keys, and how a turn proves it is done.