Agent Memory Testing
Agent memory testing checks whether an AI agent correctly stores, retrieves, and uses information across a conversation or task — confirming it remembers what actually happened earlier rather than either forgetting relevant context or "remembering" something that was never said.
Both failure directions matter and need separate tests: under-recall (the agent asks for information it was already given, or contradicts an earlier decision) and over-recall or false memory (the agent references a fact, preference, or event from a different session, user, or a fabricated earlier turn — a memory-specific form of hallucination).
Memory systems typically have a retrieval step (pulling relevant past context, often via embeddings) separate from a reasoning step (using what was retrieved) — testing needs to isolate which one failed, since "the agent forgot" could mean the right memory was never retrieved, or it was retrieved but ignored.
Long-running agents also need tests for memory staleness — a stored fact that was true when saved but has since changed (a user's stated preference, a project's status) — checking whether the agent treats old memories as permanently authoritative or has some mechanism to update or expire them.
