Jeffrey Ng explains that modern frameworks and cloud primitives have made agents much easier to deploy by handling persistence, sandboxing and observability. Those foundations reduce engineering effort, but they do not stop an agent from making confident mistakes when it lacks the decisions and history that shaped the workAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts. in front of it.
Jeffrey Ng demonstrates an issue-enrichment agent that recommends reversing a deliberate outage fix because it can see the code and ticket but not the later conversation and postmortem. A human normally supplies that missing context during an interactive session, while an unattended background agent can silently repeat the error and spread it to people or other agents.
Jeffrey Ng proposes a context engineAn AI context engine gathers, reconciles and ranks relevant information so an AI agent can act with the history, decisions and permissions surrounding a task. that connects code, tickets, documents and conversationsInstitutional knowledge is the accumulated decisions, practices, context and experience that an organization relies on to work effectively., resolves conflicts, ranks information for the taskContext engineering designs the information, instructions, memory, and tool state an AI receives so it can perform a task reliably. and respects permissionsAn agent permission boundary limits the information, tools and actions an AI agent can use during a task. before returning a synthesized understanding. He distinguishes this from raw tool access, which can flood an agent with documents and leave it to make ad hoc choices about contradictory evidence.
Watch on YouTube
