What problem does it solve? Designing memory and state for an AI agent system is easy to get wrong: teams either over-engineer storage they don't need or forget durable state and audit trails that later stages depend on. This Skill walks through a six-category memory taxonomy and produces an approved, written memory design consistent with earlier architecture decisions. ## Core Features & Use Cases - Six-category memory taxonomy: Evaluates conversation context, user preferences, task/scratch state, business records, long-term knowledge, and audit history, deciding for each whether it is needed and how it is stored and scoped. - Cross-stage consistency checks: Aligns task-state durability with the runtime shape from stage 3 and audit history with the tool inventory from stage 4, flagging inconsistencies instead of silently re-deciding earlier choices. - Approval-gated output: Writes a structured system_design/05_memory.md decision table and stops for explicit user approval before the pipeline continues. - Use Case: As stage 5 of an agent-system-design pipeline, after topology, pattern, runtime shape, and tools are approved, use it to decide that a support agent needs per-session conversation context, checkpointed task state for human approvals, and an append-only audit log, while skipping user preferences for anonymous usage. ## Quick Start Run stage 5 of the agent system design pipeline to decide memory and state storage for my approved agent design and write the memory design document for my review.
