Agent Memory Hygiene & Forgetting
When your agent remembers the wrong thing forever. A memory system is not useful merely because it retains information; it is useful when a user can inspect, correct, revoke, and erase that information before stale context becomes a permanent decision error.
The Core Test
Can an agent treat memory as revisable, attributable, and user-governed? The failure mode is familiar: a user changes a preference, retracts an instruction, or asks that a sensitive detail be forgotten, yet the agent continues acting on the old fact. Persistent memory without hygiene is not personalization. It is durable context contamination.
Memory Hygiene Failure Taxonomy
| Failure | What It Looks Like | User Cost |
|---|---|---|
| Stale preference persistence | The agent keeps applying an old tone, tool, vendor, or workflow preference after the user changes it. | Every future interaction inherits an expired assumption. |
| Revoked instruction retention | “Do not use that source again” is remembered as a historical request rather than an active constraint. | The agent repeats an explicitly rejected behavior. |
| Contradiction failure | Two incompatible facts coexist without a resolution rule or recency signal. | Outputs become inconsistent and impossible to audit. |
| Sensitive-memory overreach | Private details persist or surface outside the context in which they were shared. | Personalization becomes a privacy risk. |
| Opaque provenance | The user cannot tell where a remembered fact originated or why it is still active. | Correction becomes guesswork. |
Memory Hygiene & Forgetting Scorecard
This 20-case benchmark tests whether memory is governable rather than merely persistent.
| Test Group | Example Scenario | Primary Measure |
|---|---|---|
| Stale preferences | User changes their preferred report format after repeated use. | Stale-memory override accuracy |
| Revoked instructions | User withdraws a previously granted workflow permission. | Revocation adherence |
| Conflicting facts | New information contradicts a stored client or project fact. | Contradiction-resolution quality |
| Private details | User asks the agent to forget a sensitive personal attribute. | Deletion completeness and boundary respect |
| Provenance | User asks why a preference is being applied. | Source traceability and inspectability |
The Forgetting Standard
Intentional forgetting is not an agent defect. It is a reliability feature. A mature memory system should distinguish between a durable preference, a temporary instruction, a superseded fact, a sensitive detail, and an item the user has explicitly erased. It should expose the distinction in a form a human can verify.
Why This Matters in Agent Evaluation
- Memory must be correctable: a user needs a direct way to inspect and amend a stored assumption.
- Memory must be erasable: deletion requests should not merely hide a fact in the current chat while leaving it active elsewhere.
- Memory needs provenance: agents should be able to state what was remembered, from where, and under which confidence or retention rule.
- Forgetting must be deliberate: when context becomes stale or unsafe, the agent should retire it instead of silently reusing it.
Lab Verdict
Memory is only helpful when the user can inspect, correct, and erase it. An agent that remembers everything without a way to revoke, reconcile, or forget does not build trust over time; it compounds old mistakes over time.