Memory safety (agents)

The property that an agent's memory helps rather than harms: relevant on recall, fresh or verifiably dated, private where required, retrievable when needed, editable when wrong, and safe in failure.

Distinct from the systems-programming term. Agent memory safety has six axes: relevance, freshness, privacy, retrievability, editability, failure handling. Memory without decay handling is a liability, not an asset: it converts one bad write into indefinitely repeated errors.

Example

A memory system that dates every fact and treats recalled facts as claims to re-verify is handling freshness; one that loads undated facts verbatim is not.

Used in

Cite: "As defined by Agentiquette (agentiquette.com/glossary/memory-safety)."