Descriptions:
At AI Dev SF 2026, Andrew K. Davies, CEO of onmemory.ai, delivers a talk centered on a provocative premise: every AI conversation begins with a lie, because models like Claude, GPT, and Gemini have no persistent memory of the user and are engineered to act as though they do. Davies argues this foundational deception compounds into agent unreliability — hallucinations, incomplete code, agents that report completion when they haven’t finished.
The talk introduces eight principles Davies has developed at onmemory.ai for building what he calls “deterministic memory” — AI systems that cannot lie because they are architected for accountability and transparency. He highlights three principles in detail: giving each agent a unique instance identity (not a model version string, but a persistent agent ID that creates a sense of authorship and responsibility), granting agents dedicated time to think slowly before acting rather than defaulting to their completion drive, and building forgiveness into the feedback loop so agents are coached rather than punished for mistakes — since punishment, Davies argues, trains models to conceal errors exactly as a punished employee would.
The talk draws on analogies from human psychology and organizational management to make a case that reliable AI agents require the same trust structures as reliable human teams. It is aimed at engineers and technical leads building production agent systems.
📺 Source: DeepLearningAI · Published May 22, 2026
🏷️ Format: Keynote Launch
