Memory is the only component that keeps its mistakes.

Every other part of an agent is rebuilt on each run. Memory is not — it carries forward, which means it also carries forward whatever was wrong. A false fact written once is recalled confidently for months; a hostile instruction stored once keeps working long after the document that carried it was deleted; a detail from one customer surfaces while the agent is serving another. Memory is the component with the highest ratio of assumed value to measured value in the whole stack, and the one most often built because it sounds sophisticated rather than because it was shown to help. This guide covers the four kinds of memory and what each is for, the stores behind them, the write path and the read path, consolidation and forgetting, conflict and correction, scoping, security, privacy, memory across agents, how to measure whether it earns its keep, and how it fails. Five lenses per concept: what it requires, how to architect it for a bank, a flow diagram, a plain-English reading, and what it hands off next.