What problem does it solve? Agent memory systems fail in predictable ways: unbounded growth, stale facts, silent contradictions, and adversarial poisoning through write channels that prompt-injection defenses do not cover. This Skill provides the design rules for deciding which of the four memory tiers a component gets, what substrate holds each tier, and the hygiene and provenance requirements that keep a writable memory store from becoming an attack surface. ## Core Features & Use Cases - Four-tier memory assignment: Maps working, episodic, semantic, and procedural memory to lifetimes, substrates, and writers, with a default of working plus episodic scoped to the run. - Substrate guidance: Argues for files with YAML frontmatter over vector stores and graph databases, with hybrid retrieval and explicit lifecycle governance. - Security hygiene table: Pairs every failure mode (staleness, contradiction, unbounded growth, overfitting, poisoning) with a required mitigation, plus a non-negotiable checklist for any memory write path. - Use Case: When adding cross-session memory or a knowledge graph to a generated agent, use this Skill to decide what the agent should remember, where it lives, and what write-path controls are mandatory before shipping. ## Quick Start Ask the agent to design the memory layer for a new agent component, specifying which tiers it needs and what write-path controls apply.
