Kay Malcolm describes a distributed software team whose AI coding tools made individuals faster without improving the team's overall throughput. Code moved between regions, but the intent, decisions, and session context behind it did not, leaving later developers to reconstruct what happened while repositories diverged.
An enterprise agent is more than a model and workflow. It also needs tools, context, memory, retrieval, security, and guardrails. In this framing, the harness gives the model ways to act, while memory carries context across the system and retrieval selects what should return for the current task.
The talk separates memory into short-term session state, long-term information across sessions, episodic records of prior interactions, procedural knowledge about tools and steps, and semantic knowledge about the organization. These forms have different access patterns, but they still need a coherent source of truth.
Scattering relational, document, graph, and vector data across isolated systems can force an agent to reconcile competing answers while consuming more tokens. A shared durable memory layer can instead preserve facts, decisions, and process history across agents and sessions while keeping retrieval and access controls explicit.
In the team example, a memory broker retained context alongside branches and commits so work could pass between people without losing the reasoning behind it. The broader lesson is that shared memory, not code generation alone, turns faster individuals into a coordinated human-agent team while leaving developers in control.
Watch on YouTube
