AI Automation ·
Deleting a conversation can leave agent memories behind: how should Korean support services verify completion?
AWS announced general availability of OpenSearch Memory in a November 18, 2026 launch article. It derives facts, preferences and summaries from agent interactions for retrieval across sessions.[1] The question raised by this news is less how long an agent can remember than what a service can promise when a customer asks it to forget. This discussion is based on public documents; we have not tested deletion using customer data.
What should a deletion button promise?
Consider the joint decision facing a Korean online service's support manager, privacy officer and engineering team. A customer withdraws an earlier preference or personal detail from future support use. If the visible transcript disappears but a summary enters the next conversation, the interface and the system disagree. Removing all context, however, may also disrupt a continuing case. This is an adoption design problem, not a report of an incident at a named Korean company.
The research question is specific: does a successful deletion response for an original event establish that related derived memories will not influence the next support conversation? Stopping use, deleting within the service and managing records subject to separate retention requirements are different decisions. We do not claim that every record must always be physically destroyed immediately.
The documentation separates two lifecycles
Initially it was unclear whether a conversation and its summary formed one deletion unit. The DeleteEvent reference answers directly: deleting an event does not delete derived memories because a memory may combine several events; removing memories requires DeleteMemory.[2] The answer to the original question is therefore no. Original-event deletion alone cannot establish cessation of downstream use.
The architecture guide supplies the next distinction. Memories record source event IDs; event and memory retention are configured separately, and memories may remain indefinitely without a memory expiry setting. Search results can briefly persist after a deletion call returns.[3] DeleteMemory removes the memory's content and embedding.[4] These are documented product behaviors, not a defect discovered by our investigation. This review did not establish a fixed maximum search delay, deletion of every application cache, or physical erasure times covering backups.
Source IDs provide a starting point for tracing information, not proof that every relevant derivative has been identified. Consolidated summaries, directly written application notes and separately retained support caches need examination in the actual data flow. The risk pathway is an inference: a surviving summary could reintroduce outdated information as present context. We have no measured Korean incident frequency or loss estimate.
Dates also need care. The launch article is dated November 18, while the developer-document history records an introduction entry on November 16.[1][6] These dates describe different publication channels; they do not establish the first operational moment in every region. The technical documents come from the same supplier and are not independent performance validation.
| Decision-relevant item | Overseas evidence | Verified Korean reference |
|---|---|---|
| Date and evidence type | November 18 product article and public API reference[1][2] | PIPC guidance announcement, August 6, 2025[5] |
| Scope to manage | Separate operations for original events and derived memories[2][4] | Lifecycle safeguards and consideration of individual rights[5] |
| Responsibility boundary | Event and memory expiry configured separately[3] | Governance centered on the privacy officer throughout the lifecycle[5] |
The PIPC announcement addresses generative AI development and use, agent management and individual rights during application.[5] It does not approve this product's Korean deployment or document a local trial. The comparison changes the implementation proposal: before treating a cloud API response as completion, support and privacy owners must define deletion, retained records and cessation of use, while engineers trace information from input to response. Actual contracts and processing purposes require separate assessment.
Limited memory or no persistent personal memory?
Option A retains only a limited set of customer-confirmed fields, with source IDs and change history. Rather than accepting every inferred memory, customers could inspect and correct explicit items such as contact preferences. On a request, an application-controlled retrieval barrier would first prevent the information from entering responses. The service would then identify the relevant originals and derivatives, perform appropriate deletion and verify the result. This barrier is an engineering proposal, not a feature we verified as automatically supplied.
A is worth considering where continuity has substantial value and the team can maintain data lineage. Confirmation screens, lineage records, retries and failure handling cost money and staff time. Removing consolidated summaries may also discard valid context. API provision does not transfer responsibility for interpreting a customer request or notifying the customer to the vendor. We propose withholding a completion notice when prevention of reintroduction has not been verified.
Option B uses the current session and approved general knowledge without persistent personal memory. It reduces management of derived personal records and stale preferences, but may increase repeated explanations and staff lookup time. Disabling long-term memory does not eliminate other logs or data stores. B is a reasonable starting point where the value of continuity remains unproven or lineage cannot be maintained. The reviewed sources do not support a monetary estimate of which option is cheaper.
Testing absence rather than an API response
Before using real personal information, we propose a test environment with fictional customers, preferences and multiple sessions. Privacy and support owners define the scope of cessation and any separate retention justification. Engineers map originals, derivatives and retrieval caches. Tests should include several memories from one event and one memory consolidated from several events. Cross-customer isolation needs a separate test.
Use the same support scenarios for a no-memory baseline B and limited-memory treatment A. Record repeated questions, resolution accuracy, storage/extraction/deletion handling costs, time from request to retrieval blocking, and the rate at which withdrawn information returns in later responses. Requery over a declared observation period and again after background processing. These are proposed measurements, not reported AWS or Korean results.
If another customer's information is returned or withdrawn information re-enters a response, we propose stopping the trial and disabling the affected memory use. Zero observed recurrence cannot prove erasure in every future condition; sample size and observation period must accompany the result. If support benefits do not appear, narrow A further or return to B. Verify that separately retained records are excluded from the relevant retrieval route.
A narrower answer and unresolved layers
The investigation resolves one issue: originals and derived memories have different lifecycles. A Korean support service should therefore not promise cessation of future use based solely on a transcript-deletion response. This finding neither prohibits every deployment nor proves that our proposed barrier is complete.
Evidence that could change the implementation judgment includes reproducible tests identifying the requested derivatives and controlling reintroduction, failure and latency behavior, and measured support benefits. Another open question is how much customer inspection and correction reduce the burden. The central debate is whether forgetting can be verified as an operational function separate from claims about an agent's ability to remember.