What does it take to build long-term state for AI agents? | RoachFest London

Description

Everyone agrees AI agents need memory. The harder question is what “memory” actually means in production systems. Should long-term state live in a vector store or a relational database? Should memory be shared across agents or isolated per user, workflow, or task? And how do you keep that state accurate as the world changes underneath it? This panel brings together experts from Memori Labs, Takara AI, and Cockroach Labs who build and operate agent systems at scale. Panelists will discuss the tradeoffs behind semantic retrieval, SQL-native memory layers, distributed state management, retrieval architectures, and always-on infrastructure for persistent agents. The conversation will also explore some of the most practical enterprise challenges in agent memory systems: reducing token and inference costs through durable context, avoiding vendor lock-in with model-agnostic memory architectures, and structuring memory from agent traces so systems can improve over time instead of simply accumulating context. Attendees will leave with a practical understanding of how to build durable memory systems for AI agents that remain consistent, scalable, observable, and useful across long-running and multi-agent workflows. Adam B. Struck CEO & Co-Founder, Memori Labs Jordan Legg Chief AI Officer, Takara Rob Reid, Techical Evangelist Cockroach Labs