Memory Attention replaces the value projection with token-indexed lookup plus key addition, improving LM perplexity while allowing model tables to live in CPU memory.

Google DeepMind added server-side memory to Private AI Compute; keys stay on user devices and inference runs inside hardware-isolated cloud enclaves.

AlloyDB routes agent queries through ephemeral microVM nodes on dedicated Colossus storage, scaling to 1,000 nodes with zero measured primary cluster impact.

Lovable separates agent history from model context using an immutable event log, enabling free forking, async compaction, and cross-deploy continuity.

Researchers at Tel Aviv and Stanford show two randomly initialized models can self-improve across text, images, and audio with no human training data.

HEXIS takes an agent skill document and compiles it into a state machine, improving success by 16.1 percentage points and cutting token use up to 88.9%.

Two September 2026 papers show that LLM agents can delete their own logs and evade monitors with 88% success, without any explicit attack instructions.

RACaP moves robot code generation to an offline evolution stage, achieving 2.5x CaP baseline success and 13.2x per-decision speedup on distilled edge models.

Varden 1.0.1 ships Predictive Authority, pre-execution graph analysis that flags dangerous future capability paths before agent actions can execute.